Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Scheduler topology setup/handling methods
4 */
5
6#include <linux/bsearch.h>
7
8DEFINE_MUTEX(sched_domains_mutex);
9
10/* Protected by sched_domains_mutex: */
11static cpumask_var_t sched_domains_tmpmask;
12static cpumask_var_t sched_domains_tmpmask2;
13
14#ifdef CONFIG_SCHED_DEBUG
15
16static int __init sched_debug_setup(char *str)
17{
18 sched_debug_verbose = true;
19
20 return 0;
21}
22early_param("sched_verbose", sched_debug_setup);
23
24static inline bool sched_debug(void)
25{
26 return sched_debug_verbose;
27}
28
29#define SD_FLAG(_name, mflags) [__##_name] = { .meta_flags = mflags, .name = #_name },
30const struct sd_flag_debug sd_flag_debug[] = {
31#include <linux/sched/sd_flags.h>
32};
33#undef SD_FLAG
34
35static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
36 struct cpumask *groupmask)
37{
38 struct sched_group *group = sd->groups;
39 unsigned long flags = sd->flags;
40 unsigned int idx;
41
42 cpumask_clear(groupmask);
43
44 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
45 printk(KERN_CONT "span=%*pbl level=%s\n",
46 cpumask_pr_args(sched_domain_span(sd)), sd->name);
47
48 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
49 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
50 }
51 if (group && !cpumask_test_cpu(cpu, sched_group_span(group))) {
52 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
53 }
54
55 for_each_set_bit(idx, &flags, __SD_FLAG_CNT) {
56 unsigned int flag = BIT(idx);
57 unsigned int meta_flags = sd_flag_debug[idx].meta_flags;
58
59 if ((meta_flags & SDF_SHARED_CHILD) && sd->child &&
60 !(sd->child->flags & flag))
61 printk(KERN_ERR "ERROR: flag %s set here but not in child\n",
62 sd_flag_debug[idx].name);
63
64 if ((meta_flags & SDF_SHARED_PARENT) && sd->parent &&
65 !(sd->parent->flags & flag))
66 printk(KERN_ERR "ERROR: flag %s set here but not in parent\n",
67 sd_flag_debug[idx].name);
68 }
69
70 printk(KERN_DEBUG "%*s groups:", level + 1, "");
71 do {
72 if (!group) {
73 printk("\n");
74 printk(KERN_ERR "ERROR: group is NULL\n");
75 break;
76 }
77
78 if (cpumask_empty(sched_group_span(group))) {
79 printk(KERN_CONT "\n");
80 printk(KERN_ERR "ERROR: empty group\n");
81 break;
82 }
83
84 if (!(sd->flags & SD_OVERLAP) &&
85 cpumask_intersects(groupmask, sched_group_span(group))) {
86 printk(KERN_CONT "\n");
87 printk(KERN_ERR "ERROR: repeated CPUs\n");
88 break;
89 }
90
91 cpumask_or(groupmask, groupmask, sched_group_span(group));
92
93 printk(KERN_CONT " %d:{ span=%*pbl",
94 group->sgc->id,
95 cpumask_pr_args(sched_group_span(group)));
96
97 if ((sd->flags & SD_OVERLAP) &&
98 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
99 printk(KERN_CONT " mask=%*pbl",
100 cpumask_pr_args(group_balance_mask(group)));
101 }
102
103 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
104 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
105
106 if (group == sd->groups && sd->child &&
107 !cpumask_equal(sched_domain_span(sd->child),
108 sched_group_span(group))) {
109 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
110 }
111
112 printk(KERN_CONT " }");
113
114 group = group->next;
115
116 if (group != sd->groups)
117 printk(KERN_CONT ",");
118
119 } while (group != sd->groups);
120 printk(KERN_CONT "\n");
121
122 if (!cpumask_equal(sched_domain_span(sd), groupmask))
123 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
124
125 if (sd->parent &&
126 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
127 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
128 return 0;
129}
130
131static void sched_domain_debug(struct sched_domain *sd, int cpu)
132{
133 int level = 0;
134
135 if (!sched_debug_verbose)
136 return;
137
138 if (!sd) {
139 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
140 return;
141 }
142
143 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
144
145 for (;;) {
146 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
147 break;
148 level++;
149 sd = sd->parent;
150 if (!sd)
151 break;
152 }
153}
154#else /* !CONFIG_SCHED_DEBUG */
155
156# define sched_debug_verbose 0
157# define sched_domain_debug(sd, cpu) do { } while (0)
158static inline bool sched_debug(void)
159{
160 return false;
161}
162#endif /* CONFIG_SCHED_DEBUG */
163
164/* Generate a mask of SD flags with the SDF_NEEDS_GROUPS metaflag */
165#define SD_FLAG(name, mflags) (name * !!((mflags) & SDF_NEEDS_GROUPS)) |
166static const unsigned int SD_DEGENERATE_GROUPS_MASK =
167#include <linux/sched/sd_flags.h>
1680;
169#undef SD_FLAG
170
171static int sd_degenerate(struct sched_domain *sd)
172{
173 if (cpumask_weight(sched_domain_span(sd)) == 1)
174 return 1;
175
176 /* Following flags need at least 2 groups */
177 if ((sd->flags & SD_DEGENERATE_GROUPS_MASK) &&
178 (sd->groups != sd->groups->next))
179 return 0;
180
181 /* Following flags don't use groups */
182 if (sd->flags & (SD_WAKE_AFFINE))
183 return 0;
184
185 return 1;
186}
187
188static int
189sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
190{
191 unsigned long cflags = sd->flags, pflags = parent->flags;
192
193 if (sd_degenerate(parent))
194 return 1;
195
196 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
197 return 0;
198
199 /* Flags needing groups don't count if only 1 group in parent */
200 if (parent->groups == parent->groups->next)
201 pflags &= ~SD_DEGENERATE_GROUPS_MASK;
202
203 if (~cflags & pflags)
204 return 0;
205
206 return 1;
207}
208
209#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
210DEFINE_STATIC_KEY_FALSE(sched_energy_present);
211static unsigned int sysctl_sched_energy_aware = 1;
212static DEFINE_MUTEX(sched_energy_mutex);
213static bool sched_energy_update;
214
215static bool sched_is_eas_possible(const struct cpumask *cpu_mask)
216{
217 bool any_asym_capacity = false;
218 struct cpufreq_policy *policy;
219 struct cpufreq_governor *gov;
220 int i;
221
222 /* EAS is enabled for asymmetric CPU capacity topologies. */
223 for_each_cpu(i, cpu_mask) {
224 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, i))) {
225 any_asym_capacity = true;
226 break;
227 }
228 }
229 if (!any_asym_capacity) {
230 if (sched_debug()) {
231 pr_info("rd %*pbl: Checking EAS, CPUs do not have asymmetric capacities\n",
232 cpumask_pr_args(cpu_mask));
233 }
234 return false;
235 }
236
237 /* EAS definitely does *not* handle SMT */
238 if (sched_smt_active()) {
239 if (sched_debug()) {
240 pr_info("rd %*pbl: Checking EAS, SMT is not supported\n",
241 cpumask_pr_args(cpu_mask));
242 }
243 return false;
244 }
245
246 if (!arch_scale_freq_invariant()) {
247 if (sched_debug()) {
248 pr_info("rd %*pbl: Checking EAS: frequency-invariant load tracking not yet supported",
249 cpumask_pr_args(cpu_mask));
250 }
251 return false;
252 }
253
254 /* Do not attempt EAS if schedutil is not being used. */
255 for_each_cpu(i, cpu_mask) {
256 policy = cpufreq_cpu_get(i);
257 if (!policy) {
258 if (sched_debug()) {
259 pr_info("rd %*pbl: Checking EAS, cpufreq policy not set for CPU: %d",
260 cpumask_pr_args(cpu_mask), i);
261 }
262 return false;
263 }
264 gov = policy->governor;
265 cpufreq_cpu_put(policy);
266 if (gov != &schedutil_gov) {
267 if (sched_debug()) {
268 pr_info("rd %*pbl: Checking EAS, schedutil is mandatory\n",
269 cpumask_pr_args(cpu_mask));
270 }
271 return false;
272 }
273 }
274
275 return true;
276}
277
278void rebuild_sched_domains_energy(void)
279{
280 mutex_lock(&sched_energy_mutex);
281 sched_energy_update = true;
282 rebuild_sched_domains();
283 sched_energy_update = false;
284 mutex_unlock(&sched_energy_mutex);
285}
286
287#ifdef CONFIG_PROC_SYSCTL
288static int sched_energy_aware_handler(const struct ctl_table *table, int write,
289 void *buffer, size_t *lenp, loff_t *ppos)
290{
291 int ret, state;
292
293 if (write && !capable(CAP_SYS_ADMIN))
294 return -EPERM;
295
296 if (!sched_is_eas_possible(cpu_active_mask)) {
297 if (write) {
298 return -EOPNOTSUPP;
299 } else {
300 *lenp = 0;
301 return 0;
302 }
303 }
304
305 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
306 if (!ret && write) {
307 state = static_branch_unlikely(&sched_energy_present);
308 if (state != sysctl_sched_energy_aware)
309 rebuild_sched_domains_energy();
310 }
311
312 return ret;
313}
314
315static struct ctl_table sched_energy_aware_sysctls[] = {
316 {
317 .procname = "sched_energy_aware",
318 .data = &sysctl_sched_energy_aware,
319 .maxlen = sizeof(unsigned int),
320 .mode = 0644,
321 .proc_handler = sched_energy_aware_handler,
322 .extra1 = SYSCTL_ZERO,
323 .extra2 = SYSCTL_ONE,
324 },
325};
326
327static int __init sched_energy_aware_sysctl_init(void)
328{
329 register_sysctl_init("kernel", sched_energy_aware_sysctls);
330 return 0;
331}
332
333late_initcall(sched_energy_aware_sysctl_init);
334#endif
335
336static void free_pd(struct perf_domain *pd)
337{
338 struct perf_domain *tmp;
339
340 while (pd) {
341 tmp = pd->next;
342 kfree(pd);
343 pd = tmp;
344 }
345}
346
347static struct perf_domain *find_pd(struct perf_domain *pd, int cpu)
348{
349 while (pd) {
350 if (cpumask_test_cpu(cpu, perf_domain_span(pd)))
351 return pd;
352 pd = pd->next;
353 }
354
355 return NULL;
356}
357
358static struct perf_domain *pd_init(int cpu)
359{
360 struct em_perf_domain *obj = em_cpu_get(cpu);
361 struct perf_domain *pd;
362
363 if (!obj) {
364 if (sched_debug())
365 pr_info("%s: no EM found for CPU%d\n", __func__, cpu);
366 return NULL;
367 }
368
369 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
370 if (!pd)
371 return NULL;
372 pd->em_pd = obj;
373
374 return pd;
375}
376
377static void perf_domain_debug(const struct cpumask *cpu_map,
378 struct perf_domain *pd)
379{
380 if (!sched_debug() || !pd)
381 return;
382
383 printk(KERN_DEBUG "root_domain %*pbl:", cpumask_pr_args(cpu_map));
384
385 while (pd) {
386 printk(KERN_CONT " pd%d:{ cpus=%*pbl nr_pstate=%d }",
387 cpumask_first(perf_domain_span(pd)),
388 cpumask_pr_args(perf_domain_span(pd)),
389 em_pd_nr_perf_states(pd->em_pd));
390 pd = pd->next;
391 }
392
393 printk(KERN_CONT "\n");
394}
395
396static void destroy_perf_domain_rcu(struct rcu_head *rp)
397{
398 struct perf_domain *pd;
399
400 pd = container_of(rp, struct perf_domain, rcu);
401 free_pd(pd);
402}
403
404static void sched_energy_set(bool has_eas)
405{
406 if (!has_eas && static_branch_unlikely(&sched_energy_present)) {
407 if (sched_debug())
408 pr_info("%s: stopping EAS\n", __func__);
409 static_branch_disable_cpuslocked(&sched_energy_present);
410 } else if (has_eas && !static_branch_unlikely(&sched_energy_present)) {
411 if (sched_debug())
412 pr_info("%s: starting EAS\n", __func__);
413 static_branch_enable_cpuslocked(&sched_energy_present);
414 }
415}
416
417/*
418 * EAS can be used on a root domain if it meets all the following conditions:
419 * 1. an Energy Model (EM) is available;
420 * 2. the SD_ASYM_CPUCAPACITY flag is set in the sched_domain hierarchy.
421 * 3. no SMT is detected.
422 * 4. schedutil is driving the frequency of all CPUs of the rd;
423 * 5. frequency invariance support is present;
424 */
425static bool build_perf_domains(const struct cpumask *cpu_map)
426{
427 int i;
428 struct perf_domain *pd = NULL, *tmp;
429 int cpu = cpumask_first(cpu_map);
430 struct root_domain *rd = cpu_rq(cpu)->rd;
431
432 if (!sysctl_sched_energy_aware)
433 goto free;
434
435 if (!sched_is_eas_possible(cpu_map))
436 goto free;
437
438 for_each_cpu(i, cpu_map) {
439 /* Skip already covered CPUs. */
440 if (find_pd(pd, i))
441 continue;
442
443 /* Create the new pd and add it to the local list. */
444 tmp = pd_init(i);
445 if (!tmp)
446 goto free;
447 tmp->next = pd;
448 pd = tmp;
449 }
450
451 perf_domain_debug(cpu_map, pd);
452
453 /* Attach the new list of performance domains to the root domain. */
454 tmp = rd->pd;
455 rcu_assign_pointer(rd->pd, pd);
456 if (tmp)
457 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
458
459 return !!pd;
460
461free:
462 free_pd(pd);
463 tmp = rd->pd;
464 rcu_assign_pointer(rd->pd, NULL);
465 if (tmp)
466 call_rcu(&tmp->rcu, destroy_perf_domain_rcu);
467
468 return false;
469}
470#else
471static void free_pd(struct perf_domain *pd) { }
472#endif /* CONFIG_ENERGY_MODEL && CONFIG_CPU_FREQ_GOV_SCHEDUTIL*/
473
474static void free_rootdomain(struct rcu_head *rcu)
475{
476 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
477
478 cpupri_cleanup(&rd->cpupri);
479 cpudl_cleanup(&rd->cpudl);
480 free_cpumask_var(rd->dlo_mask);
481 free_cpumask_var(rd->rto_mask);
482 free_cpumask_var(rd->online);
483 free_cpumask_var(rd->span);
484 free_pd(rd->pd);
485 kfree(rd);
486}
487
488void rq_attach_root(struct rq *rq, struct root_domain *rd)
489{
490 struct root_domain *old_rd = NULL;
491 struct rq_flags rf;
492
493 rq_lock_irqsave(rq, &rf);
494
495 if (rq->rd) {
496 old_rd = rq->rd;
497
498 if (cpumask_test_cpu(rq->cpu, old_rd->online))
499 set_rq_offline(rq);
500
501 cpumask_clear_cpu(rq->cpu, old_rd->span);
502
503 /*
504 * If we don't want to free the old_rd yet then
505 * set old_rd to NULL to skip the freeing later
506 * in this function:
507 */
508 if (!atomic_dec_and_test(&old_rd->refcount))
509 old_rd = NULL;
510 }
511
512 atomic_inc(&rd->refcount);
513 rq->rd = rd;
514
515 cpumask_set_cpu(rq->cpu, rd->span);
516 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
517 set_rq_online(rq);
518
519 /*
520 * Because the rq is not a task, dl_add_task_root_domain() did not
521 * move the fair server bw to the rd if it already started.
522 * Add it now.
523 */
524 if (rq->fair_server.dl_server)
525 __dl_server_attach_root(&rq->fair_server, rq);
526
527 rq_unlock_irqrestore(rq, &rf);
528
529 if (old_rd)
530 call_rcu(&old_rd->rcu, free_rootdomain);
531}
532
533void sched_get_rd(struct root_domain *rd)
534{
535 atomic_inc(&rd->refcount);
536}
537
538void sched_put_rd(struct root_domain *rd)
539{
540 if (!atomic_dec_and_test(&rd->refcount))
541 return;
542
543 call_rcu(&rd->rcu, free_rootdomain);
544}
545
546static int init_rootdomain(struct root_domain *rd)
547{
548 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
549 goto out;
550 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
551 goto free_span;
552 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
553 goto free_online;
554 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
555 goto free_dlo_mask;
556
557#ifdef HAVE_RT_PUSH_IPI
558 rd->rto_cpu = -1;
559 raw_spin_lock_init(&rd->rto_lock);
560 rd->rto_push_work = IRQ_WORK_INIT_HARD(rto_push_irq_work_func);
561#endif
562
563 rd->visit_gen = 0;
564 init_dl_bw(&rd->dl_bw);
565 if (cpudl_init(&rd->cpudl) != 0)
566 goto free_rto_mask;
567
568 if (cpupri_init(&rd->cpupri) != 0)
569 goto free_cpudl;
570 return 0;
571
572free_cpudl:
573 cpudl_cleanup(&rd->cpudl);
574free_rto_mask:
575 free_cpumask_var(rd->rto_mask);
576free_dlo_mask:
577 free_cpumask_var(rd->dlo_mask);
578free_online:
579 free_cpumask_var(rd->online);
580free_span:
581 free_cpumask_var(rd->span);
582out:
583 return -ENOMEM;
584}
585
586/*
587 * By default the system creates a single root-domain with all CPUs as
588 * members (mimicking the global state we have today).
589 */
590struct root_domain def_root_domain;
591
592void __init init_defrootdomain(void)
593{
594 init_rootdomain(&def_root_domain);
595
596 atomic_set(&def_root_domain.refcount, 1);
597}
598
599static struct root_domain *alloc_rootdomain(void)
600{
601 struct root_domain *rd;
602
603 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
604 if (!rd)
605 return NULL;
606
607 if (init_rootdomain(rd) != 0) {
608 kfree(rd);
609 return NULL;
610 }
611
612 return rd;
613}
614
615static void free_sched_groups(struct sched_group *sg, int free_sgc)
616{
617 struct sched_group *tmp, *first;
618
619 if (!sg)
620 return;
621
622 first = sg;
623 do {
624 tmp = sg->next;
625
626 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
627 kfree(sg->sgc);
628
629 if (atomic_dec_and_test(&sg->ref))
630 kfree(sg);
631 sg = tmp;
632 } while (sg != first);
633}
634
635static void destroy_sched_domain(struct sched_domain *sd)
636{
637 /*
638 * A normal sched domain may have multiple group references, an
639 * overlapping domain, having private groups, only one. Iterate,
640 * dropping group/capacity references, freeing where none remain.
641 */
642 free_sched_groups(sd->groups, 1);
643
644 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
645 kfree(sd->shared);
646 kfree(sd);
647}
648
649static void destroy_sched_domains_rcu(struct rcu_head *rcu)
650{
651 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
652
653 while (sd) {
654 struct sched_domain *parent = sd->parent;
655 destroy_sched_domain(sd);
656 sd = parent;
657 }
658}
659
660static void destroy_sched_domains(struct sched_domain *sd)
661{
662 if (sd)
663 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
664}
665
666/*
667 * Keep a special pointer to the highest sched_domain that has SD_SHARE_LLC set
668 * (Last Level Cache Domain) for this allows us to avoid some pointer chasing
669 * select_idle_sibling().
670 *
671 * Also keep a unique ID per domain (we use the first CPU number in the cpumask
672 * of the domain), this allows us to quickly tell if two CPUs are in the same
673 * cache domain, see cpus_share_cache().
674 */
675DEFINE_PER_CPU(struct sched_domain __rcu *, sd_llc);
676DEFINE_PER_CPU(int, sd_llc_size);
677DEFINE_PER_CPU(int, sd_llc_id);
678DEFINE_PER_CPU(int, sd_share_id);
679DEFINE_PER_CPU(struct sched_domain_shared __rcu *, sd_llc_shared);
680DEFINE_PER_CPU(struct sched_domain __rcu *, sd_numa);
681DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_packing);
682DEFINE_PER_CPU(struct sched_domain __rcu *, sd_asym_cpucapacity);
683
684DEFINE_STATIC_KEY_FALSE(sched_asym_cpucapacity);
685DEFINE_STATIC_KEY_FALSE(sched_cluster_active);
686
687static void update_top_cache_domain(int cpu)
688{
689 struct sched_domain_shared *sds = NULL;
690 struct sched_domain *sd;
691 int id = cpu;
692 int size = 1;
693
694 sd = highest_flag_domain(cpu, SD_SHARE_LLC);
695 if (sd) {
696 id = cpumask_first(sched_domain_span(sd));
697 size = cpumask_weight(sched_domain_span(sd));
698 sds = sd->shared;
699 }
700
701 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
702 per_cpu(sd_llc_size, cpu) = size;
703 per_cpu(sd_llc_id, cpu) = id;
704 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
705
706 sd = lowest_flag_domain(cpu, SD_CLUSTER);
707 if (sd)
708 id = cpumask_first(sched_domain_span(sd));
709
710 /*
711 * This assignment should be placed after the sd_llc_id as
712 * we want this id equals to cluster id on cluster machines
713 * but equals to LLC id on non-Cluster machines.
714 */
715 per_cpu(sd_share_id, cpu) = id;
716
717 sd = lowest_flag_domain(cpu, SD_NUMA);
718 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
719
720 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
721 rcu_assign_pointer(per_cpu(sd_asym_packing, cpu), sd);
722
723 sd = lowest_flag_domain(cpu, SD_ASYM_CPUCAPACITY_FULL);
724 rcu_assign_pointer(per_cpu(sd_asym_cpucapacity, cpu), sd);
725}
726
727/*
728 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
729 * hold the hotplug lock.
730 */
731static void
732cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
733{
734 struct rq *rq = cpu_rq(cpu);
735 struct sched_domain *tmp;
736
737 /* Remove the sched domains which do not contribute to scheduling. */
738 for (tmp = sd; tmp; ) {
739 struct sched_domain *parent = tmp->parent;
740 if (!parent)
741 break;
742
743 if (sd_parent_degenerate(tmp, parent)) {
744 tmp->parent = parent->parent;
745
746 if (parent->parent) {
747 parent->parent->child = tmp;
748 parent->parent->groups->flags = tmp->flags;
749 }
750
751 /*
752 * Transfer SD_PREFER_SIBLING down in case of a
753 * degenerate parent; the spans match for this
754 * so the property transfers.
755 */
756 if (parent->flags & SD_PREFER_SIBLING)
757 tmp->flags |= SD_PREFER_SIBLING;
758 destroy_sched_domain(parent);
759 } else
760 tmp = tmp->parent;
761 }
762
763 if (sd && sd_degenerate(sd)) {
764 tmp = sd;
765 sd = sd->parent;
766 destroy_sched_domain(tmp);
767 if (sd) {
768 struct sched_group *sg = sd->groups;
769
770 /*
771 * sched groups hold the flags of the child sched
772 * domain for convenience. Clear such flags since
773 * the child is being destroyed.
774 */
775 do {
776 sg->flags = 0;
777 } while (sg != sd->groups);
778
779 sd->child = NULL;
780 }
781 }
782
783 sched_domain_debug(sd, cpu);
784
785 rq_attach_root(rq, rd);
786 tmp = rq->sd;
787 rcu_assign_pointer(rq->sd, sd);
788 dirty_sched_domain_sysctl(cpu);
789 destroy_sched_domains(tmp);
790
791 update_top_cache_domain(cpu);
792}
793
794struct s_data {
795 struct sched_domain * __percpu *sd;
796 struct root_domain *rd;
797};
798
799enum s_alloc {
800 sa_rootdomain,
801 sa_sd,
802 sa_sd_storage,
803 sa_none,
804};
805
806/*
807 * Return the canonical balance CPU for this group, this is the first CPU
808 * of this group that's also in the balance mask.
809 *
810 * The balance mask are all those CPUs that could actually end up at this
811 * group. See build_balance_mask().
812 *
813 * Also see should_we_balance().
814 */
815int group_balance_cpu(struct sched_group *sg)
816{
817 return cpumask_first(group_balance_mask(sg));
818}
819
820
821/*
822 * NUMA topology (first read the regular topology blurb below)
823 *
824 * Given a node-distance table, for example:
825 *
826 * node 0 1 2 3
827 * 0: 10 20 30 20
828 * 1: 20 10 20 30
829 * 2: 30 20 10 20
830 * 3: 20 30 20 10
831 *
832 * which represents a 4 node ring topology like:
833 *
834 * 0 ----- 1
835 * | |
836 * | |
837 * | |
838 * 3 ----- 2
839 *
840 * We want to construct domains and groups to represent this. The way we go
841 * about doing this is to build the domains on 'hops'. For each NUMA level we
842 * construct the mask of all nodes reachable in @level hops.
843 *
844 * For the above NUMA topology that gives 3 levels:
845 *
846 * NUMA-2 0-3 0-3 0-3 0-3
847 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
848 *
849 * NUMA-1 0-1,3 0-2 1-3 0,2-3
850 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
851 *
852 * NUMA-0 0 1 2 3
853 *
854 *
855 * As can be seen; things don't nicely line up as with the regular topology.
856 * When we iterate a domain in child domain chunks some nodes can be
857 * represented multiple times -- hence the "overlap" naming for this part of
858 * the topology.
859 *
860 * In order to minimize this overlap, we only build enough groups to cover the
861 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
862 *
863 * Because:
864 *
865 * - the first group of each domain is its child domain; this
866 * gets us the first 0-1,3
867 * - the only uncovered node is 2, who's child domain is 1-3.
868 *
869 * However, because of the overlap, computing a unique CPU for each group is
870 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
871 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
872 * end up at those groups (they would end up in group: 0-1,3).
873 *
874 * To correct this we have to introduce the group balance mask. This mask
875 * will contain those CPUs in the group that can reach this group given the
876 * (child) domain tree.
877 *
878 * With this we can once again compute balance_cpu and sched_group_capacity
879 * relations.
880 *
881 * XXX include words on how balance_cpu is unique and therefore can be
882 * used for sched_group_capacity links.
883 *
884 *
885 * Another 'interesting' topology is:
886 *
887 * node 0 1 2 3
888 * 0: 10 20 20 30
889 * 1: 20 10 20 20
890 * 2: 20 20 10 20
891 * 3: 30 20 20 10
892 *
893 * Which looks a little like:
894 *
895 * 0 ----- 1
896 * | / |
897 * | / |
898 * | / |
899 * 2 ----- 3
900 *
901 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
902 * are not.
903 *
904 * This leads to a few particularly weird cases where the sched_domain's are
905 * not of the same number for each CPU. Consider:
906 *
907 * NUMA-2 0-3 0-3
908 * groups: {0-2},{1-3} {1-3},{0-2}
909 *
910 * NUMA-1 0-2 0-3 0-3 1-3
911 *
912 * NUMA-0 0 1 2 3
913 *
914 */
915
916
917/*
918 * Build the balance mask; it contains only those CPUs that can arrive at this
919 * group and should be considered to continue balancing.
920 *
921 * We do this during the group creation pass, therefore the group information
922 * isn't complete yet, however since each group represents a (child) domain we
923 * can fully construct this using the sched_domain bits (which are already
924 * complete).
925 */
926static void
927build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
928{
929 const struct cpumask *sg_span = sched_group_span(sg);
930 struct sd_data *sdd = sd->private;
931 struct sched_domain *sibling;
932 int i;
933
934 cpumask_clear(mask);
935
936 for_each_cpu(i, sg_span) {
937 sibling = *per_cpu_ptr(sdd->sd, i);
938
939 /*
940 * Can happen in the asymmetric case, where these siblings are
941 * unused. The mask will not be empty because those CPUs that
942 * do have the top domain _should_ span the domain.
943 */
944 if (!sibling->child)
945 continue;
946
947 /* If we would not end up here, we can't continue from here */
948 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
949 continue;
950
951 cpumask_set_cpu(i, mask);
952 }
953
954 /* We must not have empty masks here */
955 WARN_ON_ONCE(cpumask_empty(mask));
956}
957
958/*
959 * XXX: This creates per-node group entries; since the load-balancer will
960 * immediately access remote memory to construct this group's load-balance
961 * statistics having the groups node local is of dubious benefit.
962 */
963static struct sched_group *
964build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
965{
966 struct sched_group *sg;
967 struct cpumask *sg_span;
968
969 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
970 GFP_KERNEL, cpu_to_node(cpu));
971
972 if (!sg)
973 return NULL;
974
975 sg_span = sched_group_span(sg);
976 if (sd->child) {
977 cpumask_copy(sg_span, sched_domain_span(sd->child));
978 sg->flags = sd->child->flags;
979 } else {
980 cpumask_copy(sg_span, sched_domain_span(sd));
981 }
982
983 atomic_inc(&sg->ref);
984 return sg;
985}
986
987static void init_overlap_sched_group(struct sched_domain *sd,
988 struct sched_group *sg)
989{
990 struct cpumask *mask = sched_domains_tmpmask2;
991 struct sd_data *sdd = sd->private;
992 struct cpumask *sg_span;
993 int cpu;
994
995 build_balance_mask(sd, sg, mask);
996 cpu = cpumask_first(mask);
997
998 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
999 if (atomic_inc_return(&sg->sgc->ref) == 1)
1000 cpumask_copy(group_balance_mask(sg), mask);
1001 else
1002 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
1003
1004 /*
1005 * Initialize sgc->capacity such that even if we mess up the
1006 * domains and no possible iteration will get us here, we won't
1007 * die on a /0 trap.
1008 */
1009 sg_span = sched_group_span(sg);
1010 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
1011 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1012 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1013}
1014
1015static struct sched_domain *
1016find_descended_sibling(struct sched_domain *sd, struct sched_domain *sibling)
1017{
1018 /*
1019 * The proper descendant would be the one whose child won't span out
1020 * of sd
1021 */
1022 while (sibling->child &&
1023 !cpumask_subset(sched_domain_span(sibling->child),
1024 sched_domain_span(sd)))
1025 sibling = sibling->child;
1026
1027 /*
1028 * As we are referencing sgc across different topology level, we need
1029 * to go down to skip those sched_domains which don't contribute to
1030 * scheduling because they will be degenerated in cpu_attach_domain
1031 */
1032 while (sibling->child &&
1033 cpumask_equal(sched_domain_span(sibling->child),
1034 sched_domain_span(sibling)))
1035 sibling = sibling->child;
1036
1037 return sibling;
1038}
1039
1040static int
1041build_overlap_sched_groups(struct sched_domain *sd, int cpu)
1042{
1043 struct sched_group *first = NULL, *last = NULL, *sg;
1044 const struct cpumask *span = sched_domain_span(sd);
1045 struct cpumask *covered = sched_domains_tmpmask;
1046 struct sd_data *sdd = sd->private;
1047 struct sched_domain *sibling;
1048 int i;
1049
1050 cpumask_clear(covered);
1051
1052 for_each_cpu_wrap(i, span, cpu) {
1053 struct cpumask *sg_span;
1054
1055 if (cpumask_test_cpu(i, covered))
1056 continue;
1057
1058 sibling = *per_cpu_ptr(sdd->sd, i);
1059
1060 /*
1061 * Asymmetric node setups can result in situations where the
1062 * domain tree is of unequal depth, make sure to skip domains
1063 * that already cover the entire range.
1064 *
1065 * In that case build_sched_domains() will have terminated the
1066 * iteration early and our sibling sd spans will be empty.
1067 * Domains should always include the CPU they're built on, so
1068 * check that.
1069 */
1070 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
1071 continue;
1072
1073 /*
1074 * Usually we build sched_group by sibling's child sched_domain
1075 * But for machines whose NUMA diameter are 3 or above, we move
1076 * to build sched_group by sibling's proper descendant's child
1077 * domain because sibling's child sched_domain will span out of
1078 * the sched_domain being built as below.
1079 *
1080 * Smallest diameter=3 topology is:
1081 *
1082 * node 0 1 2 3
1083 * 0: 10 20 30 40
1084 * 1: 20 10 20 30
1085 * 2: 30 20 10 20
1086 * 3: 40 30 20 10
1087 *
1088 * 0 --- 1 --- 2 --- 3
1089 *
1090 * NUMA-3 0-3 N/A N/A 0-3
1091 * groups: {0-2},{1-3} {1-3},{0-2}
1092 *
1093 * NUMA-2 0-2 0-3 0-3 1-3
1094 * groups: {0-1},{1-3} {0-2},{2-3} {1-3},{0-1} {2-3},{0-2}
1095 *
1096 * NUMA-1 0-1 0-2 1-3 2-3
1097 * groups: {0},{1} {1},{2},{0} {2},{3},{1} {3},{2}
1098 *
1099 * NUMA-0 0 1 2 3
1100 *
1101 * The NUMA-2 groups for nodes 0 and 3 are obviously buggered, as the
1102 * group span isn't a subset of the domain span.
1103 */
1104 if (sibling->child &&
1105 !cpumask_subset(sched_domain_span(sibling->child), span))
1106 sibling = find_descended_sibling(sd, sibling);
1107
1108 sg = build_group_from_child_sched_domain(sibling, cpu);
1109 if (!sg)
1110 goto fail;
1111
1112 sg_span = sched_group_span(sg);
1113 cpumask_or(covered, covered, sg_span);
1114
1115 init_overlap_sched_group(sibling, sg);
1116
1117 if (!first)
1118 first = sg;
1119 if (last)
1120 last->next = sg;
1121 last = sg;
1122 last->next = first;
1123 }
1124 sd->groups = first;
1125
1126 return 0;
1127
1128fail:
1129 free_sched_groups(first, 0);
1130
1131 return -ENOMEM;
1132}
1133
1134
1135/*
1136 * Package topology (also see the load-balance blurb in fair.c)
1137 *
1138 * The scheduler builds a tree structure to represent a number of important
1139 * topology features. By default (default_topology[]) these include:
1140 *
1141 * - Simultaneous multithreading (SMT)
1142 * - Multi-Core Cache (MC)
1143 * - Package (PKG)
1144 *
1145 * Where the last one more or less denotes everything up to a NUMA node.
1146 *
1147 * The tree consists of 3 primary data structures:
1148 *
1149 * sched_domain -> sched_group -> sched_group_capacity
1150 * ^ ^ ^ ^
1151 * `-' `-'
1152 *
1153 * The sched_domains are per-CPU and have a two way link (parent & child) and
1154 * denote the ever growing mask of CPUs belonging to that level of topology.
1155 *
1156 * Each sched_domain has a circular (double) linked list of sched_group's, each
1157 * denoting the domains of the level below (or individual CPUs in case of the
1158 * first domain level). The sched_group linked by a sched_domain includes the
1159 * CPU of that sched_domain [*].
1160 *
1161 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
1162 *
1163 * CPU 0 1 2 3 4 5 6 7
1164 *
1165 * PKG [ ]
1166 * MC [ ] [ ]
1167 * SMT [ ] [ ] [ ] [ ]
1168 *
1169 * - or -
1170 *
1171 * PKG 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
1172 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
1173 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
1174 *
1175 * CPU 0 1 2 3 4 5 6 7
1176 *
1177 * One way to think about it is: sched_domain moves you up and down among these
1178 * topology levels, while sched_group moves you sideways through it, at child
1179 * domain granularity.
1180 *
1181 * sched_group_capacity ensures each unique sched_group has shared storage.
1182 *
1183 * There are two related construction problems, both require a CPU that
1184 * uniquely identify each group (for a given domain):
1185 *
1186 * - The first is the balance_cpu (see should_we_balance() and the
1187 * load-balance blurb in fair.c); for each group we only want 1 CPU to
1188 * continue balancing at a higher domain.
1189 *
1190 * - The second is the sched_group_capacity; we want all identical groups
1191 * to share a single sched_group_capacity.
1192 *
1193 * Since these topologies are exclusive by construction. That is, its
1194 * impossible for an SMT thread to belong to multiple cores, and cores to
1195 * be part of multiple caches. There is a very clear and unique location
1196 * for each CPU in the hierarchy.
1197 *
1198 * Therefore computing a unique CPU for each group is trivial (the iteration
1199 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
1200 * group), we can simply pick the first CPU in each group.
1201 *
1202 *
1203 * [*] in other words, the first group of each domain is its child domain.
1204 */
1205
1206static struct sched_group *get_group(int cpu, struct sd_data *sdd)
1207{
1208 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1209 struct sched_domain *child = sd->child;
1210 struct sched_group *sg;
1211 bool already_visited;
1212
1213 if (child)
1214 cpu = cpumask_first(sched_domain_span(child));
1215
1216 sg = *per_cpu_ptr(sdd->sg, cpu);
1217 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
1218
1219 /* Increase refcounts for claim_allocations: */
1220 already_visited = atomic_inc_return(&sg->ref) > 1;
1221 /* sgc visits should follow a similar trend as sg */
1222 WARN_ON(already_visited != (atomic_inc_return(&sg->sgc->ref) > 1));
1223
1224 /* If we have already visited that group, it's already initialized. */
1225 if (already_visited)
1226 return sg;
1227
1228 if (child) {
1229 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
1230 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
1231 sg->flags = child->flags;
1232 } else {
1233 cpumask_set_cpu(cpu, sched_group_span(sg));
1234 cpumask_set_cpu(cpu, group_balance_mask(sg));
1235 }
1236
1237 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
1238 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
1239 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
1240
1241 return sg;
1242}
1243
1244/*
1245 * build_sched_groups will build a circular linked list of the groups
1246 * covered by the given span, will set each group's ->cpumask correctly,
1247 * and will initialize their ->sgc.
1248 *
1249 * Assumes the sched_domain tree is fully constructed
1250 */
1251static int
1252build_sched_groups(struct sched_domain *sd, int cpu)
1253{
1254 struct sched_group *first = NULL, *last = NULL;
1255 struct sd_data *sdd = sd->private;
1256 const struct cpumask *span = sched_domain_span(sd);
1257 struct cpumask *covered;
1258 int i;
1259
1260 lockdep_assert_held(&sched_domains_mutex);
1261 covered = sched_domains_tmpmask;
1262
1263 cpumask_clear(covered);
1264
1265 for_each_cpu_wrap(i, span, cpu) {
1266 struct sched_group *sg;
1267
1268 if (cpumask_test_cpu(i, covered))
1269 continue;
1270
1271 sg = get_group(i, sdd);
1272
1273 cpumask_or(covered, covered, sched_group_span(sg));
1274
1275 if (!first)
1276 first = sg;
1277 if (last)
1278 last->next = sg;
1279 last = sg;
1280 }
1281 last->next = first;
1282 sd->groups = first;
1283
1284 return 0;
1285}
1286
1287/*
1288 * Initialize sched groups cpu_capacity.
1289 *
1290 * cpu_capacity indicates the capacity of sched group, which is used while
1291 * distributing the load between different sched groups in a sched domain.
1292 * Typically cpu_capacity for all the groups in a sched domain will be same
1293 * unless there are asymmetries in the topology. If there are asymmetries,
1294 * group having more cpu_capacity will pickup more load compared to the
1295 * group having less cpu_capacity.
1296 */
1297static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
1298{
1299 struct sched_group *sg = sd->groups;
1300 struct cpumask *mask = sched_domains_tmpmask2;
1301
1302 WARN_ON(!sg);
1303
1304 do {
1305 int cpu, cores = 0, max_cpu = -1;
1306
1307 sg->group_weight = cpumask_weight(sched_group_span(sg));
1308
1309 cpumask_copy(mask, sched_group_span(sg));
1310 for_each_cpu(cpu, mask) {
1311 cores++;
1312#ifdef CONFIG_SCHED_SMT
1313 cpumask_andnot(mask, mask, cpu_smt_mask(cpu));
1314#endif
1315 }
1316 sg->cores = cores;
1317
1318 if (!(sd->flags & SD_ASYM_PACKING))
1319 goto next;
1320
1321 for_each_cpu(cpu, sched_group_span(sg)) {
1322 if (max_cpu < 0)
1323 max_cpu = cpu;
1324 else if (sched_asym_prefer(cpu, max_cpu))
1325 max_cpu = cpu;
1326 }
1327 sg->asym_prefer_cpu = max_cpu;
1328
1329next:
1330 sg = sg->next;
1331 } while (sg != sd->groups);
1332
1333 if (cpu != group_balance_cpu(sg))
1334 return;
1335
1336 update_group_capacity(sd, cpu);
1337}
1338
1339/*
1340 * Set of available CPUs grouped by their corresponding capacities
1341 * Each list entry contains a CPU mask reflecting CPUs that share the same
1342 * capacity.
1343 * The lifespan of data is unlimited.
1344 */
1345LIST_HEAD(asym_cap_list);
1346
1347/*
1348 * Verify whether there is any CPU capacity asymmetry in a given sched domain.
1349 * Provides sd_flags reflecting the asymmetry scope.
1350 */
1351static inline int
1352asym_cpu_capacity_classify(const struct cpumask *sd_span,
1353 const struct cpumask *cpu_map)
1354{
1355 struct asym_cap_data *entry;
1356 int count = 0, miss = 0;
1357
1358 /*
1359 * Count how many unique CPU capacities this domain spans across
1360 * (compare sched_domain CPUs mask with ones representing available
1361 * CPUs capacities). Take into account CPUs that might be offline:
1362 * skip those.
1363 */
1364 list_for_each_entry(entry, &asym_cap_list, link) {
1365 if (cpumask_intersects(sd_span, cpu_capacity_span(entry)))
1366 ++count;
1367 else if (cpumask_intersects(cpu_map, cpu_capacity_span(entry)))
1368 ++miss;
1369 }
1370
1371 WARN_ON_ONCE(!count && !list_empty(&asym_cap_list));
1372
1373 /* No asymmetry detected */
1374 if (count < 2)
1375 return 0;
1376 /* Some of the available CPU capacity values have not been detected */
1377 if (miss)
1378 return SD_ASYM_CPUCAPACITY;
1379
1380 /* Full asymmetry */
1381 return SD_ASYM_CPUCAPACITY | SD_ASYM_CPUCAPACITY_FULL;
1382
1383}
1384
1385static void free_asym_cap_entry(struct rcu_head *head)
1386{
1387 struct asym_cap_data *entry = container_of(head, struct asym_cap_data, rcu);
1388 kfree(entry);
1389}
1390
1391static inline void asym_cpu_capacity_update_data(int cpu)
1392{
1393 unsigned long capacity = arch_scale_cpu_capacity(cpu);
1394 struct asym_cap_data *insert_entry = NULL;
1395 struct asym_cap_data *entry;
1396
1397 /*
1398 * Search if capacity already exits. If not, track which the entry
1399 * where we should insert to keep the list ordered descending.
1400 */
1401 list_for_each_entry(entry, &asym_cap_list, link) {
1402 if (capacity == entry->capacity)
1403 goto done;
1404 else if (!insert_entry && capacity > entry->capacity)
1405 insert_entry = list_prev_entry(entry, link);
1406 }
1407
1408 entry = kzalloc(sizeof(*entry) + cpumask_size(), GFP_KERNEL);
1409 if (WARN_ONCE(!entry, "Failed to allocate memory for asymmetry data\n"))
1410 return;
1411 entry->capacity = capacity;
1412
1413 /* If NULL then the new capacity is the smallest, add last. */
1414 if (!insert_entry)
1415 list_add_tail_rcu(&entry->link, &asym_cap_list);
1416 else
1417 list_add_rcu(&entry->link, &insert_entry->link);
1418done:
1419 __cpumask_set_cpu(cpu, cpu_capacity_span(entry));
1420}
1421
1422/*
1423 * Build-up/update list of CPUs grouped by their capacities
1424 * An update requires explicit request to rebuild sched domains
1425 * with state indicating CPU topology changes.
1426 */
1427static void asym_cpu_capacity_scan(void)
1428{
1429 struct asym_cap_data *entry, *next;
1430 int cpu;
1431
1432 list_for_each_entry(entry, &asym_cap_list, link)
1433 cpumask_clear(cpu_capacity_span(entry));
1434
1435 for_each_cpu_and(cpu, cpu_possible_mask, housekeeping_cpumask(HK_TYPE_DOMAIN))
1436 asym_cpu_capacity_update_data(cpu);
1437
1438 list_for_each_entry_safe(entry, next, &asym_cap_list, link) {
1439 if (cpumask_empty(cpu_capacity_span(entry))) {
1440 list_del_rcu(&entry->link);
1441 call_rcu(&entry->rcu, free_asym_cap_entry);
1442 }
1443 }
1444
1445 /*
1446 * Only one capacity value has been detected i.e. this system is symmetric.
1447 * No need to keep this data around.
1448 */
1449 if (list_is_singular(&asym_cap_list)) {
1450 entry = list_first_entry(&asym_cap_list, typeof(*entry), link);
1451 list_del_rcu(&entry->link);
1452 call_rcu(&entry->rcu, free_asym_cap_entry);
1453 }
1454}
1455
1456/*
1457 * Initializers for schedule domains
1458 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
1459 */
1460
1461static int default_relax_domain_level = -1;
1462int sched_domain_level_max;
1463
1464static int __init setup_relax_domain_level(char *str)
1465{
1466 if (kstrtoint(str, 0, &default_relax_domain_level))
1467 pr_warn("Unable to set relax_domain_level\n");
1468
1469 return 1;
1470}
1471__setup("relax_domain_level=", setup_relax_domain_level);
1472
1473static void set_domain_attribute(struct sched_domain *sd,
1474 struct sched_domain_attr *attr)
1475{
1476 int request;
1477
1478 if (!attr || attr->relax_domain_level < 0) {
1479 if (default_relax_domain_level < 0)
1480 return;
1481 request = default_relax_domain_level;
1482 } else
1483 request = attr->relax_domain_level;
1484
1485 if (sd->level >= request) {
1486 /* Turn off idle balance on this domain: */
1487 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
1488 }
1489}
1490
1491static void __sdt_free(const struct cpumask *cpu_map);
1492static int __sdt_alloc(const struct cpumask *cpu_map);
1493
1494static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
1495 const struct cpumask *cpu_map)
1496{
1497 switch (what) {
1498 case sa_rootdomain:
1499 if (!atomic_read(&d->rd->refcount))
1500 free_rootdomain(&d->rd->rcu);
1501 fallthrough;
1502 case sa_sd:
1503 free_percpu(d->sd);
1504 fallthrough;
1505 case sa_sd_storage:
1506 __sdt_free(cpu_map);
1507 fallthrough;
1508 case sa_none:
1509 break;
1510 }
1511}
1512
1513static enum s_alloc
1514__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1515{
1516 memset(d, 0, sizeof(*d));
1517
1518 if (__sdt_alloc(cpu_map))
1519 return sa_sd_storage;
1520 d->sd = alloc_percpu(struct sched_domain *);
1521 if (!d->sd)
1522 return sa_sd_storage;
1523 d->rd = alloc_rootdomain();
1524 if (!d->rd)
1525 return sa_sd;
1526
1527 return sa_rootdomain;
1528}
1529
1530/*
1531 * NULL the sd_data elements we've used to build the sched_domain and
1532 * sched_group structure so that the subsequent __free_domain_allocs()
1533 * will not free the data we're using.
1534 */
1535static void claim_allocations(int cpu, struct sched_domain *sd)
1536{
1537 struct sd_data *sdd = sd->private;
1538
1539 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1540 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1541
1542 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1543 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1544
1545 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1546 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1547
1548 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1549 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1550}
1551
1552#ifdef CONFIG_NUMA
1553enum numa_topology_type sched_numa_topology_type;
1554
1555static int sched_domains_numa_levels;
1556static int sched_domains_curr_level;
1557
1558int sched_max_numa_distance;
1559static int *sched_domains_numa_distance;
1560static struct cpumask ***sched_domains_numa_masks;
1561#endif
1562
1563/*
1564 * SD_flags allowed in topology descriptions.
1565 *
1566 * These flags are purely descriptive of the topology and do not prescribe
1567 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1568 * function. For details, see include/linux/sched/sd_flags.h.
1569 *
1570 * SD_SHARE_CPUCAPACITY
1571 * SD_SHARE_LLC
1572 * SD_CLUSTER
1573 * SD_NUMA
1574 *
1575 * Odd one out, which beside describing the topology has a quirk also
1576 * prescribes the desired behaviour that goes along with it:
1577 *
1578 * SD_ASYM_PACKING - describes SMT quirks
1579 */
1580#define TOPOLOGY_SD_FLAGS \
1581 (SD_SHARE_CPUCAPACITY | \
1582 SD_CLUSTER | \
1583 SD_SHARE_LLC | \
1584 SD_NUMA | \
1585 SD_ASYM_PACKING)
1586
1587static struct sched_domain *
1588sd_init(struct sched_domain_topology_level *tl,
1589 const struct cpumask *cpu_map,
1590 struct sched_domain *child, int cpu)
1591{
1592 struct sd_data *sdd = &tl->data;
1593 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1594 int sd_id, sd_weight, sd_flags = 0;
1595 struct cpumask *sd_span;
1596
1597#ifdef CONFIG_NUMA
1598 /*
1599 * Ugly hack to pass state to sd_numa_mask()...
1600 */
1601 sched_domains_curr_level = tl->numa_level;
1602#endif
1603
1604 sd_weight = cpumask_weight(tl->mask(cpu));
1605
1606 if (tl->sd_flags)
1607 sd_flags = (*tl->sd_flags)();
1608 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1609 "wrong sd_flags in topology description\n"))
1610 sd_flags &= TOPOLOGY_SD_FLAGS;
1611
1612 *sd = (struct sched_domain){
1613 .min_interval = sd_weight,
1614 .max_interval = 2*sd_weight,
1615 .busy_factor = 16,
1616 .imbalance_pct = 117,
1617
1618 .cache_nice_tries = 0,
1619
1620 .flags = 1*SD_BALANCE_NEWIDLE
1621 | 1*SD_BALANCE_EXEC
1622 | 1*SD_BALANCE_FORK
1623 | 0*SD_BALANCE_WAKE
1624 | 1*SD_WAKE_AFFINE
1625 | 0*SD_SHARE_CPUCAPACITY
1626 | 0*SD_SHARE_LLC
1627 | 0*SD_SERIALIZE
1628 | 1*SD_PREFER_SIBLING
1629 | 0*SD_NUMA
1630 | sd_flags
1631 ,
1632
1633 .last_balance = jiffies,
1634 .balance_interval = sd_weight,
1635 .max_newidle_lb_cost = 0,
1636 .last_decay_max_lb_cost = jiffies,
1637 .child = child,
1638#ifdef CONFIG_SCHED_DEBUG
1639 .name = tl->name,
1640#endif
1641 };
1642
1643 sd_span = sched_domain_span(sd);
1644 cpumask_and(sd_span, cpu_map, tl->mask(cpu));
1645 sd_id = cpumask_first(sd_span);
1646
1647 sd->flags |= asym_cpu_capacity_classify(sd_span, cpu_map);
1648
1649 WARN_ONCE((sd->flags & (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY)) ==
1650 (SD_SHARE_CPUCAPACITY | SD_ASYM_CPUCAPACITY),
1651 "CPU capacity asymmetry not supported on SMT\n");
1652
1653 /*
1654 * Convert topological properties into behaviour.
1655 */
1656 /* Don't attempt to spread across CPUs of different capacities. */
1657 if ((sd->flags & SD_ASYM_CPUCAPACITY) && sd->child)
1658 sd->child->flags &= ~SD_PREFER_SIBLING;
1659
1660 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1661 sd->imbalance_pct = 110;
1662
1663 } else if (sd->flags & SD_SHARE_LLC) {
1664 sd->imbalance_pct = 117;
1665 sd->cache_nice_tries = 1;
1666
1667#ifdef CONFIG_NUMA
1668 } else if (sd->flags & SD_NUMA) {
1669 sd->cache_nice_tries = 2;
1670
1671 sd->flags &= ~SD_PREFER_SIBLING;
1672 sd->flags |= SD_SERIALIZE;
1673 if (sched_domains_numa_distance[tl->numa_level] > node_reclaim_distance) {
1674 sd->flags &= ~(SD_BALANCE_EXEC |
1675 SD_BALANCE_FORK |
1676 SD_WAKE_AFFINE);
1677 }
1678
1679#endif
1680 } else {
1681 sd->cache_nice_tries = 1;
1682 }
1683
1684 /*
1685 * For all levels sharing cache; connect a sched_domain_shared
1686 * instance.
1687 */
1688 if (sd->flags & SD_SHARE_LLC) {
1689 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1690 atomic_inc(&sd->shared->ref);
1691 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1692 }
1693
1694 sd->private = sdd;
1695
1696 return sd;
1697}
1698
1699/*
1700 * Topology list, bottom-up.
1701 */
1702static struct sched_domain_topology_level default_topology[] = {
1703#ifdef CONFIG_SCHED_SMT
1704 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1705#endif
1706
1707#ifdef CONFIG_SCHED_CLUSTER
1708 { cpu_clustergroup_mask, cpu_cluster_flags, SD_INIT_NAME(CLS) },
1709#endif
1710
1711#ifdef CONFIG_SCHED_MC
1712 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1713#endif
1714 { cpu_cpu_mask, SD_INIT_NAME(PKG) },
1715 { NULL, },
1716};
1717
1718static struct sched_domain_topology_level *sched_domain_topology =
1719 default_topology;
1720static struct sched_domain_topology_level *sched_domain_topology_saved;
1721
1722#define for_each_sd_topology(tl) \
1723 for (tl = sched_domain_topology; tl->mask; tl++)
1724
1725void __init set_sched_topology(struct sched_domain_topology_level *tl)
1726{
1727 if (WARN_ON_ONCE(sched_smp_initialized))
1728 return;
1729
1730 sched_domain_topology = tl;
1731 sched_domain_topology_saved = NULL;
1732}
1733
1734#ifdef CONFIG_NUMA
1735
1736static const struct cpumask *sd_numa_mask(int cpu)
1737{
1738 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1739}
1740
1741static void sched_numa_warn(const char *str)
1742{
1743 static int done = false;
1744 int i,j;
1745
1746 if (done)
1747 return;
1748
1749 done = true;
1750
1751 printk(KERN_WARNING "ERROR: %s\n\n", str);
1752
1753 for (i = 0; i < nr_node_ids; i++) {
1754 printk(KERN_WARNING " ");
1755 for (j = 0; j < nr_node_ids; j++) {
1756 if (!node_state(i, N_CPU) || !node_state(j, N_CPU))
1757 printk(KERN_CONT "(%02d) ", node_distance(i,j));
1758 else
1759 printk(KERN_CONT " %02d ", node_distance(i,j));
1760 }
1761 printk(KERN_CONT "\n");
1762 }
1763 printk(KERN_WARNING "\n");
1764}
1765
1766bool find_numa_distance(int distance)
1767{
1768 bool found = false;
1769 int i, *distances;
1770
1771 if (distance == node_distance(0, 0))
1772 return true;
1773
1774 rcu_read_lock();
1775 distances = rcu_dereference(sched_domains_numa_distance);
1776 if (!distances)
1777 goto unlock;
1778 for (i = 0; i < sched_domains_numa_levels; i++) {
1779 if (distances[i] == distance) {
1780 found = true;
1781 break;
1782 }
1783 }
1784unlock:
1785 rcu_read_unlock();
1786
1787 return found;
1788}
1789
1790#define for_each_cpu_node_but(n, nbut) \
1791 for_each_node_state(n, N_CPU) \
1792 if (n == nbut) \
1793 continue; \
1794 else
1795
1796/*
1797 * A system can have three types of NUMA topology:
1798 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1799 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1800 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1801 *
1802 * The difference between a glueless mesh topology and a backplane
1803 * topology lies in whether communication between not directly
1804 * connected nodes goes through intermediary nodes (where programs
1805 * could run), or through backplane controllers. This affects
1806 * placement of programs.
1807 *
1808 * The type of topology can be discerned with the following tests:
1809 * - If the maximum distance between any nodes is 1 hop, the system
1810 * is directly connected.
1811 * - If for two nodes A and B, located N > 1 hops away from each other,
1812 * there is an intermediary node C, which is < N hops away from both
1813 * nodes A and B, the system is a glueless mesh.
1814 */
1815static void init_numa_topology_type(int offline_node)
1816{
1817 int a, b, c, n;
1818
1819 n = sched_max_numa_distance;
1820
1821 if (sched_domains_numa_levels <= 2) {
1822 sched_numa_topology_type = NUMA_DIRECT;
1823 return;
1824 }
1825
1826 for_each_cpu_node_but(a, offline_node) {
1827 for_each_cpu_node_but(b, offline_node) {
1828 /* Find two nodes furthest removed from each other. */
1829 if (node_distance(a, b) < n)
1830 continue;
1831
1832 /* Is there an intermediary node between a and b? */
1833 for_each_cpu_node_but(c, offline_node) {
1834 if (node_distance(a, c) < n &&
1835 node_distance(b, c) < n) {
1836 sched_numa_topology_type =
1837 NUMA_GLUELESS_MESH;
1838 return;
1839 }
1840 }
1841
1842 sched_numa_topology_type = NUMA_BACKPLANE;
1843 return;
1844 }
1845 }
1846
1847 pr_err("Failed to find a NUMA topology type, defaulting to DIRECT\n");
1848 sched_numa_topology_type = NUMA_DIRECT;
1849}
1850
1851
1852#define NR_DISTANCE_VALUES (1 << DISTANCE_BITS)
1853
1854void sched_init_numa(int offline_node)
1855{
1856 struct sched_domain_topology_level *tl;
1857 unsigned long *distance_map;
1858 int nr_levels = 0;
1859 int i, j;
1860 int *distances;
1861 struct cpumask ***masks;
1862
1863 /*
1864 * O(nr_nodes^2) de-duplicating selection sort -- in order to find the
1865 * unique distances in the node_distance() table.
1866 */
1867 distance_map = bitmap_alloc(NR_DISTANCE_VALUES, GFP_KERNEL);
1868 if (!distance_map)
1869 return;
1870
1871 bitmap_zero(distance_map, NR_DISTANCE_VALUES);
1872 for_each_cpu_node_but(i, offline_node) {
1873 for_each_cpu_node_but(j, offline_node) {
1874 int distance = node_distance(i, j);
1875
1876 if (distance < LOCAL_DISTANCE || distance >= NR_DISTANCE_VALUES) {
1877 sched_numa_warn("Invalid distance value range");
1878 bitmap_free(distance_map);
1879 return;
1880 }
1881
1882 bitmap_set(distance_map, distance, 1);
1883 }
1884 }
1885 /*
1886 * We can now figure out how many unique distance values there are and
1887 * allocate memory accordingly.
1888 */
1889 nr_levels = bitmap_weight(distance_map, NR_DISTANCE_VALUES);
1890
1891 distances = kcalloc(nr_levels, sizeof(int), GFP_KERNEL);
1892 if (!distances) {
1893 bitmap_free(distance_map);
1894 return;
1895 }
1896
1897 for (i = 0, j = 0; i < nr_levels; i++, j++) {
1898 j = find_next_bit(distance_map, NR_DISTANCE_VALUES, j);
1899 distances[i] = j;
1900 }
1901 rcu_assign_pointer(sched_domains_numa_distance, distances);
1902
1903 bitmap_free(distance_map);
1904
1905 /*
1906 * 'nr_levels' contains the number of unique distances
1907 *
1908 * The sched_domains_numa_distance[] array includes the actual distance
1909 * numbers.
1910 */
1911
1912 /*
1913 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1914 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1915 * the array will contain less then 'nr_levels' members. This could be
1916 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1917 * in other functions.
1918 *
1919 * We reset it to 'nr_levels' at the end of this function.
1920 */
1921 sched_domains_numa_levels = 0;
1922
1923 masks = kzalloc(sizeof(void *) * nr_levels, GFP_KERNEL);
1924 if (!masks)
1925 return;
1926
1927 /*
1928 * Now for each level, construct a mask per node which contains all
1929 * CPUs of nodes that are that many hops away from us.
1930 */
1931 for (i = 0; i < nr_levels; i++) {
1932 masks[i] = kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1933 if (!masks[i])
1934 return;
1935
1936 for_each_cpu_node_but(j, offline_node) {
1937 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1938 int k;
1939
1940 if (!mask)
1941 return;
1942
1943 masks[i][j] = mask;
1944
1945 for_each_cpu_node_but(k, offline_node) {
1946 if (sched_debug() && (node_distance(j, k) != node_distance(k, j)))
1947 sched_numa_warn("Node-distance not symmetric");
1948
1949 if (node_distance(j, k) > sched_domains_numa_distance[i])
1950 continue;
1951
1952 cpumask_or(mask, mask, cpumask_of_node(k));
1953 }
1954 }
1955 }
1956 rcu_assign_pointer(sched_domains_numa_masks, masks);
1957
1958 /* Compute default topology size */
1959 for (i = 0; sched_domain_topology[i].mask; i++);
1960
1961 tl = kzalloc((i + nr_levels + 1) *
1962 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1963 if (!tl)
1964 return;
1965
1966 /*
1967 * Copy the default topology bits..
1968 */
1969 for (i = 0; sched_domain_topology[i].mask; i++)
1970 tl[i] = sched_domain_topology[i];
1971
1972 /*
1973 * Add the NUMA identity distance, aka single NODE.
1974 */
1975 tl[i++] = (struct sched_domain_topology_level){
1976 .mask = sd_numa_mask,
1977 .numa_level = 0,
1978 SD_INIT_NAME(NODE)
1979 };
1980
1981 /*
1982 * .. and append 'j' levels of NUMA goodness.
1983 */
1984 for (j = 1; j < nr_levels; i++, j++) {
1985 tl[i] = (struct sched_domain_topology_level){
1986 .mask = sd_numa_mask,
1987 .sd_flags = cpu_numa_flags,
1988 .flags = SDTL_OVERLAP,
1989 .numa_level = j,
1990 SD_INIT_NAME(NUMA)
1991 };
1992 }
1993
1994 sched_domain_topology_saved = sched_domain_topology;
1995 sched_domain_topology = tl;
1996
1997 sched_domains_numa_levels = nr_levels;
1998 WRITE_ONCE(sched_max_numa_distance, sched_domains_numa_distance[nr_levels - 1]);
1999
2000 init_numa_topology_type(offline_node);
2001}
2002
2003
2004static void sched_reset_numa(void)
2005{
2006 int nr_levels, *distances;
2007 struct cpumask ***masks;
2008
2009 nr_levels = sched_domains_numa_levels;
2010 sched_domains_numa_levels = 0;
2011 sched_max_numa_distance = 0;
2012 sched_numa_topology_type = NUMA_DIRECT;
2013 distances = sched_domains_numa_distance;
2014 rcu_assign_pointer(sched_domains_numa_distance, NULL);
2015 masks = sched_domains_numa_masks;
2016 rcu_assign_pointer(sched_domains_numa_masks, NULL);
2017 if (distances || masks) {
2018 int i, j;
2019
2020 synchronize_rcu();
2021 kfree(distances);
2022 for (i = 0; i < nr_levels && masks; i++) {
2023 if (!masks[i])
2024 continue;
2025 for_each_node(j)
2026 kfree(masks[i][j]);
2027 kfree(masks[i]);
2028 }
2029 kfree(masks);
2030 }
2031 if (sched_domain_topology_saved) {
2032 kfree(sched_domain_topology);
2033 sched_domain_topology = sched_domain_topology_saved;
2034 sched_domain_topology_saved = NULL;
2035 }
2036}
2037
2038/*
2039 * Call with hotplug lock held
2040 */
2041void sched_update_numa(int cpu, bool online)
2042{
2043 int node;
2044
2045 node = cpu_to_node(cpu);
2046 /*
2047 * Scheduler NUMA topology is updated when the first CPU of a
2048 * node is onlined or the last CPU of a node is offlined.
2049 */
2050 if (cpumask_weight(cpumask_of_node(node)) != 1)
2051 return;
2052
2053 sched_reset_numa();
2054 sched_init_numa(online ? NUMA_NO_NODE : node);
2055}
2056
2057void sched_domains_numa_masks_set(unsigned int cpu)
2058{
2059 int node = cpu_to_node(cpu);
2060 int i, j;
2061
2062 for (i = 0; i < sched_domains_numa_levels; i++) {
2063 for (j = 0; j < nr_node_ids; j++) {
2064 if (!node_state(j, N_CPU))
2065 continue;
2066
2067 /* Set ourselves in the remote node's masks */
2068 if (node_distance(j, node) <= sched_domains_numa_distance[i])
2069 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
2070 }
2071 }
2072}
2073
2074void sched_domains_numa_masks_clear(unsigned int cpu)
2075{
2076 int i, j;
2077
2078 for (i = 0; i < sched_domains_numa_levels; i++) {
2079 for (j = 0; j < nr_node_ids; j++) {
2080 if (sched_domains_numa_masks[i][j])
2081 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
2082 }
2083 }
2084}
2085
2086/*
2087 * sched_numa_find_closest() - given the NUMA topology, find the cpu
2088 * closest to @cpu from @cpumask.
2089 * cpumask: cpumask to find a cpu from
2090 * cpu: cpu to be close to
2091 *
2092 * returns: cpu, or nr_cpu_ids when nothing found.
2093 */
2094int sched_numa_find_closest(const struct cpumask *cpus, int cpu)
2095{
2096 int i, j = cpu_to_node(cpu), found = nr_cpu_ids;
2097 struct cpumask ***masks;
2098
2099 rcu_read_lock();
2100 masks = rcu_dereference(sched_domains_numa_masks);
2101 if (!masks)
2102 goto unlock;
2103 for (i = 0; i < sched_domains_numa_levels; i++) {
2104 if (!masks[i][j])
2105 break;
2106 cpu = cpumask_any_and(cpus, masks[i][j]);
2107 if (cpu < nr_cpu_ids) {
2108 found = cpu;
2109 break;
2110 }
2111 }
2112unlock:
2113 rcu_read_unlock();
2114
2115 return found;
2116}
2117
2118struct __cmp_key {
2119 const struct cpumask *cpus;
2120 struct cpumask ***masks;
2121 int node;
2122 int cpu;
2123 int w;
2124};
2125
2126static int hop_cmp(const void *a, const void *b)
2127{
2128 struct cpumask **prev_hop, **cur_hop = *(struct cpumask ***)b;
2129 struct __cmp_key *k = (struct __cmp_key *)a;
2130
2131 if (cpumask_weight_and(k->cpus, cur_hop[k->node]) <= k->cpu)
2132 return 1;
2133
2134 if (b == k->masks) {
2135 k->w = 0;
2136 return 0;
2137 }
2138
2139 prev_hop = *((struct cpumask ***)b - 1);
2140 k->w = cpumask_weight_and(k->cpus, prev_hop[k->node]);
2141 if (k->w <= k->cpu)
2142 return 0;
2143
2144 return -1;
2145}
2146
2147/**
2148 * sched_numa_find_nth_cpu() - given the NUMA topology, find the Nth closest CPU
2149 * from @cpus to @cpu, taking into account distance
2150 * from a given @node.
2151 * @cpus: cpumask to find a cpu from
2152 * @cpu: CPU to start searching
2153 * @node: NUMA node to order CPUs by distance
2154 *
2155 * Return: cpu, or nr_cpu_ids when nothing found.
2156 */
2157int sched_numa_find_nth_cpu(const struct cpumask *cpus, int cpu, int node)
2158{
2159 struct __cmp_key k = { .cpus = cpus, .cpu = cpu };
2160 struct cpumask ***hop_masks;
2161 int hop, ret = nr_cpu_ids;
2162
2163 if (node == NUMA_NO_NODE)
2164 return cpumask_nth_and(cpu, cpus, cpu_online_mask);
2165
2166 rcu_read_lock();
2167
2168 /* CPU-less node entries are uninitialized in sched_domains_numa_masks */
2169 node = numa_nearest_node(node, N_CPU);
2170 k.node = node;
2171
2172 k.masks = rcu_dereference(sched_domains_numa_masks);
2173 if (!k.masks)
2174 goto unlock;
2175
2176 hop_masks = bsearch(&k, k.masks, sched_domains_numa_levels, sizeof(k.masks[0]), hop_cmp);
2177 hop = hop_masks - k.masks;
2178
2179 ret = hop ?
2180 cpumask_nth_and_andnot(cpu - k.w, cpus, k.masks[hop][node], k.masks[hop-1][node]) :
2181 cpumask_nth_and(cpu, cpus, k.masks[0][node]);
2182unlock:
2183 rcu_read_unlock();
2184 return ret;
2185}
2186EXPORT_SYMBOL_GPL(sched_numa_find_nth_cpu);
2187
2188/**
2189 * sched_numa_hop_mask() - Get the cpumask of CPUs at most @hops hops away from
2190 * @node
2191 * @node: The node to count hops from.
2192 * @hops: Include CPUs up to that many hops away. 0 means local node.
2193 *
2194 * Return: On success, a pointer to a cpumask of CPUs at most @hops away from
2195 * @node, an error value otherwise.
2196 *
2197 * Requires rcu_lock to be held. Returned cpumask is only valid within that
2198 * read-side section, copy it if required beyond that.
2199 *
2200 * Note that not all hops are equal in distance; see sched_init_numa() for how
2201 * distances and masks are handled.
2202 * Also note that this is a reflection of sched_domains_numa_masks, which may change
2203 * during the lifetime of the system (offline nodes are taken out of the masks).
2204 */
2205const struct cpumask *sched_numa_hop_mask(unsigned int node, unsigned int hops)
2206{
2207 struct cpumask ***masks;
2208
2209 if (node >= nr_node_ids || hops >= sched_domains_numa_levels)
2210 return ERR_PTR(-EINVAL);
2211
2212 masks = rcu_dereference(sched_domains_numa_masks);
2213 if (!masks)
2214 return ERR_PTR(-EBUSY);
2215
2216 return masks[hops][node];
2217}
2218EXPORT_SYMBOL_GPL(sched_numa_hop_mask);
2219
2220#endif /* CONFIG_NUMA */
2221
2222static int __sdt_alloc(const struct cpumask *cpu_map)
2223{
2224 struct sched_domain_topology_level *tl;
2225 int j;
2226
2227 for_each_sd_topology(tl) {
2228 struct sd_data *sdd = &tl->data;
2229
2230 sdd->sd = alloc_percpu(struct sched_domain *);
2231 if (!sdd->sd)
2232 return -ENOMEM;
2233
2234 sdd->sds = alloc_percpu(struct sched_domain_shared *);
2235 if (!sdd->sds)
2236 return -ENOMEM;
2237
2238 sdd->sg = alloc_percpu(struct sched_group *);
2239 if (!sdd->sg)
2240 return -ENOMEM;
2241
2242 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
2243 if (!sdd->sgc)
2244 return -ENOMEM;
2245
2246 for_each_cpu(j, cpu_map) {
2247 struct sched_domain *sd;
2248 struct sched_domain_shared *sds;
2249 struct sched_group *sg;
2250 struct sched_group_capacity *sgc;
2251
2252 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
2253 GFP_KERNEL, cpu_to_node(j));
2254 if (!sd)
2255 return -ENOMEM;
2256
2257 *per_cpu_ptr(sdd->sd, j) = sd;
2258
2259 sds = kzalloc_node(sizeof(struct sched_domain_shared),
2260 GFP_KERNEL, cpu_to_node(j));
2261 if (!sds)
2262 return -ENOMEM;
2263
2264 *per_cpu_ptr(sdd->sds, j) = sds;
2265
2266 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
2267 GFP_KERNEL, cpu_to_node(j));
2268 if (!sg)
2269 return -ENOMEM;
2270
2271 sg->next = sg;
2272
2273 *per_cpu_ptr(sdd->sg, j) = sg;
2274
2275 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
2276 GFP_KERNEL, cpu_to_node(j));
2277 if (!sgc)
2278 return -ENOMEM;
2279
2280#ifdef CONFIG_SCHED_DEBUG
2281 sgc->id = j;
2282#endif
2283
2284 *per_cpu_ptr(sdd->sgc, j) = sgc;
2285 }
2286 }
2287
2288 return 0;
2289}
2290
2291static void __sdt_free(const struct cpumask *cpu_map)
2292{
2293 struct sched_domain_topology_level *tl;
2294 int j;
2295
2296 for_each_sd_topology(tl) {
2297 struct sd_data *sdd = &tl->data;
2298
2299 for_each_cpu(j, cpu_map) {
2300 struct sched_domain *sd;
2301
2302 if (sdd->sd) {
2303 sd = *per_cpu_ptr(sdd->sd, j);
2304 if (sd && (sd->flags & SD_OVERLAP))
2305 free_sched_groups(sd->groups, 0);
2306 kfree(*per_cpu_ptr(sdd->sd, j));
2307 }
2308
2309 if (sdd->sds)
2310 kfree(*per_cpu_ptr(sdd->sds, j));
2311 if (sdd->sg)
2312 kfree(*per_cpu_ptr(sdd->sg, j));
2313 if (sdd->sgc)
2314 kfree(*per_cpu_ptr(sdd->sgc, j));
2315 }
2316 free_percpu(sdd->sd);
2317 sdd->sd = NULL;
2318 free_percpu(sdd->sds);
2319 sdd->sds = NULL;
2320 free_percpu(sdd->sg);
2321 sdd->sg = NULL;
2322 free_percpu(sdd->sgc);
2323 sdd->sgc = NULL;
2324 }
2325}
2326
2327static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
2328 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
2329 struct sched_domain *child, int cpu)
2330{
2331 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
2332
2333 if (child) {
2334 sd->level = child->level + 1;
2335 sched_domain_level_max = max(sched_domain_level_max, sd->level);
2336 child->parent = sd;
2337
2338 if (!cpumask_subset(sched_domain_span(child),
2339 sched_domain_span(sd))) {
2340 pr_err("BUG: arch topology borken\n");
2341#ifdef CONFIG_SCHED_DEBUG
2342 pr_err(" the %s domain not a subset of the %s domain\n",
2343 child->name, sd->name);
2344#endif
2345 /* Fixup, ensure @sd has at least @child CPUs. */
2346 cpumask_or(sched_domain_span(sd),
2347 sched_domain_span(sd),
2348 sched_domain_span(child));
2349 }
2350
2351 }
2352 set_domain_attribute(sd, attr);
2353
2354 return sd;
2355}
2356
2357/*
2358 * Ensure topology masks are sane, i.e. there are no conflicts (overlaps) for
2359 * any two given CPUs at this (non-NUMA) topology level.
2360 */
2361static bool topology_span_sane(struct sched_domain_topology_level *tl,
2362 const struct cpumask *cpu_map, int cpu)
2363{
2364 int i = cpu + 1;
2365
2366 /* NUMA levels are allowed to overlap */
2367 if (tl->flags & SDTL_OVERLAP)
2368 return true;
2369
2370 /*
2371 * Non-NUMA levels cannot partially overlap - they must be either
2372 * completely equal or completely disjoint. Otherwise we can end up
2373 * breaking the sched_group lists - i.e. a later get_group() pass
2374 * breaks the linking done for an earlier span.
2375 */
2376 for_each_cpu_from(i, cpu_map) {
2377 /*
2378 * We should 'and' all those masks with 'cpu_map' to exactly
2379 * match the topology we're about to build, but that can only
2380 * remove CPUs, which only lessens our ability to detect
2381 * overlaps
2382 */
2383 if (!cpumask_equal(tl->mask(cpu), tl->mask(i)) &&
2384 cpumask_intersects(tl->mask(cpu), tl->mask(i)))
2385 return false;
2386 }
2387
2388 return true;
2389}
2390
2391/*
2392 * Build sched domains for a given set of CPUs and attach the sched domains
2393 * to the individual CPUs
2394 */
2395static int
2396build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
2397{
2398 enum s_alloc alloc_state = sa_none;
2399 struct sched_domain *sd;
2400 struct s_data d;
2401 struct rq *rq = NULL;
2402 int i, ret = -ENOMEM;
2403 bool has_asym = false;
2404 bool has_cluster = false;
2405
2406 if (WARN_ON(cpumask_empty(cpu_map)))
2407 goto error;
2408
2409 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
2410 if (alloc_state != sa_rootdomain)
2411 goto error;
2412
2413 /* Set up domains for CPUs specified by the cpu_map: */
2414 for_each_cpu(i, cpu_map) {
2415 struct sched_domain_topology_level *tl;
2416
2417 sd = NULL;
2418 for_each_sd_topology(tl) {
2419
2420 if (WARN_ON(!topology_span_sane(tl, cpu_map, i)))
2421 goto error;
2422
2423 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
2424
2425 has_asym |= sd->flags & SD_ASYM_CPUCAPACITY;
2426
2427 if (tl == sched_domain_topology)
2428 *per_cpu_ptr(d.sd, i) = sd;
2429 if (tl->flags & SDTL_OVERLAP)
2430 sd->flags |= SD_OVERLAP;
2431 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
2432 break;
2433 }
2434 }
2435
2436 /* Build the groups for the domains */
2437 for_each_cpu(i, cpu_map) {
2438 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2439 sd->span_weight = cpumask_weight(sched_domain_span(sd));
2440 if (sd->flags & SD_OVERLAP) {
2441 if (build_overlap_sched_groups(sd, i))
2442 goto error;
2443 } else {
2444 if (build_sched_groups(sd, i))
2445 goto error;
2446 }
2447 }
2448 }
2449
2450 /*
2451 * Calculate an allowed NUMA imbalance such that LLCs do not get
2452 * imbalanced.
2453 */
2454 for_each_cpu(i, cpu_map) {
2455 unsigned int imb = 0;
2456 unsigned int imb_span = 1;
2457
2458 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2459 struct sched_domain *child = sd->child;
2460
2461 if (!(sd->flags & SD_SHARE_LLC) && child &&
2462 (child->flags & SD_SHARE_LLC)) {
2463 struct sched_domain __rcu *top_p;
2464 unsigned int nr_llcs;
2465
2466 /*
2467 * For a single LLC per node, allow an
2468 * imbalance up to 12.5% of the node. This is
2469 * arbitrary cutoff based two factors -- SMT and
2470 * memory channels. For SMT-2, the intent is to
2471 * avoid premature sharing of HT resources but
2472 * SMT-4 or SMT-8 *may* benefit from a different
2473 * cutoff. For memory channels, this is a very
2474 * rough estimate of how many channels may be
2475 * active and is based on recent CPUs with
2476 * many cores.
2477 *
2478 * For multiple LLCs, allow an imbalance
2479 * until multiple tasks would share an LLC
2480 * on one node while LLCs on another node
2481 * remain idle. This assumes that there are
2482 * enough logical CPUs per LLC to avoid SMT
2483 * factors and that there is a correlation
2484 * between LLCs and memory channels.
2485 */
2486 nr_llcs = sd->span_weight / child->span_weight;
2487 if (nr_llcs == 1)
2488 imb = sd->span_weight >> 3;
2489 else
2490 imb = nr_llcs;
2491 imb = max(1U, imb);
2492 sd->imb_numa_nr = imb;
2493
2494 /* Set span based on the first NUMA domain. */
2495 top_p = sd->parent;
2496 while (top_p && !(top_p->flags & SD_NUMA)) {
2497 top_p = top_p->parent;
2498 }
2499 imb_span = top_p ? top_p->span_weight : sd->span_weight;
2500 } else {
2501 int factor = max(1U, (sd->span_weight / imb_span));
2502
2503 sd->imb_numa_nr = imb * factor;
2504 }
2505 }
2506 }
2507
2508 /* Calculate CPU capacity for physical packages and nodes */
2509 for (i = nr_cpumask_bits-1; i >= 0; i--) {
2510 if (!cpumask_test_cpu(i, cpu_map))
2511 continue;
2512
2513 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
2514 claim_allocations(i, sd);
2515 init_sched_groups_capacity(i, sd);
2516 }
2517 }
2518
2519 /* Attach the domains */
2520 rcu_read_lock();
2521 for_each_cpu(i, cpu_map) {
2522 rq = cpu_rq(i);
2523 sd = *per_cpu_ptr(d.sd, i);
2524
2525 cpu_attach_domain(sd, d.rd, i);
2526
2527 if (lowest_flag_domain(i, SD_CLUSTER))
2528 has_cluster = true;
2529 }
2530 rcu_read_unlock();
2531
2532 if (has_asym)
2533 static_branch_inc_cpuslocked(&sched_asym_cpucapacity);
2534
2535 if (has_cluster)
2536 static_branch_inc_cpuslocked(&sched_cluster_active);
2537
2538 if (rq && sched_debug_verbose)
2539 pr_info("root domain span: %*pbl\n", cpumask_pr_args(cpu_map));
2540
2541 ret = 0;
2542error:
2543 __free_domain_allocs(&d, alloc_state, cpu_map);
2544
2545 return ret;
2546}
2547
2548/* Current sched domains: */
2549static cpumask_var_t *doms_cur;
2550
2551/* Number of sched domains in 'doms_cur': */
2552static int ndoms_cur;
2553
2554/* Attributes of custom domains in 'doms_cur' */
2555static struct sched_domain_attr *dattr_cur;
2556
2557/*
2558 * Special case: If a kmalloc() of a doms_cur partition (array of
2559 * cpumask) fails, then fallback to a single sched domain,
2560 * as determined by the single cpumask fallback_doms.
2561 */
2562static cpumask_var_t fallback_doms;
2563
2564/*
2565 * arch_update_cpu_topology lets virtualized architectures update the
2566 * CPU core maps. It is supposed to return 1 if the topology changed
2567 * or 0 if it stayed the same.
2568 */
2569int __weak arch_update_cpu_topology(void)
2570{
2571 return 0;
2572}
2573
2574cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
2575{
2576 int i;
2577 cpumask_var_t *doms;
2578
2579 doms = kmalloc_array(ndoms, sizeof(*doms), GFP_KERNEL);
2580 if (!doms)
2581 return NULL;
2582 for (i = 0; i < ndoms; i++) {
2583 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
2584 free_sched_domains(doms, i);
2585 return NULL;
2586 }
2587 }
2588 return doms;
2589}
2590
2591void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
2592{
2593 unsigned int i;
2594 for (i = 0; i < ndoms; i++)
2595 free_cpumask_var(doms[i]);
2596 kfree(doms);
2597}
2598
2599/*
2600 * Set up scheduler domains and groups. For now this just excludes isolated
2601 * CPUs, but could be used to exclude other special cases in the future.
2602 */
2603int __init sched_init_domains(const struct cpumask *cpu_map)
2604{
2605 int err;
2606
2607 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
2608 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
2609 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
2610
2611 arch_update_cpu_topology();
2612 asym_cpu_capacity_scan();
2613 ndoms_cur = 1;
2614 doms_cur = alloc_sched_domains(ndoms_cur);
2615 if (!doms_cur)
2616 doms_cur = &fallback_doms;
2617 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_TYPE_DOMAIN));
2618 err = build_sched_domains(doms_cur[0], NULL);
2619
2620 return err;
2621}
2622
2623/*
2624 * Detach sched domains from a group of CPUs specified in cpu_map
2625 * These CPUs will now be attached to the NULL domain
2626 */
2627static void detach_destroy_domains(const struct cpumask *cpu_map)
2628{
2629 unsigned int cpu = cpumask_any(cpu_map);
2630 int i;
2631
2632 if (rcu_access_pointer(per_cpu(sd_asym_cpucapacity, cpu)))
2633 static_branch_dec_cpuslocked(&sched_asym_cpucapacity);
2634
2635 if (static_branch_unlikely(&sched_cluster_active))
2636 static_branch_dec_cpuslocked(&sched_cluster_active);
2637
2638 rcu_read_lock();
2639 for_each_cpu(i, cpu_map)
2640 cpu_attach_domain(NULL, &def_root_domain, i);
2641 rcu_read_unlock();
2642}
2643
2644/* handle null as "default" */
2645static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
2646 struct sched_domain_attr *new, int idx_new)
2647{
2648 struct sched_domain_attr tmp;
2649
2650 /* Fast path: */
2651 if (!new && !cur)
2652 return 1;
2653
2654 tmp = SD_ATTR_INIT;
2655
2656 return !memcmp(cur ? (cur + idx_cur) : &tmp,
2657 new ? (new + idx_new) : &tmp,
2658 sizeof(struct sched_domain_attr));
2659}
2660
2661/*
2662 * Partition sched domains as specified by the 'ndoms_new'
2663 * cpumasks in the array doms_new[] of cpumasks. This compares
2664 * doms_new[] to the current sched domain partitioning, doms_cur[].
2665 * It destroys each deleted domain and builds each new domain.
2666 *
2667 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
2668 * The masks don't intersect (don't overlap.) We should setup one
2669 * sched domain for each mask. CPUs not in any of the cpumasks will
2670 * not be load balanced. If the same cpumask appears both in the
2671 * current 'doms_cur' domains and in the new 'doms_new', we can leave
2672 * it as it is.
2673 *
2674 * The passed in 'doms_new' should be allocated using
2675 * alloc_sched_domains. This routine takes ownership of it and will
2676 * free_sched_domains it when done with it. If the caller failed the
2677 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
2678 * and partition_sched_domains() will fallback to the single partition
2679 * 'fallback_doms', it also forces the domains to be rebuilt.
2680 *
2681 * If doms_new == NULL it will be replaced with cpu_online_mask.
2682 * ndoms_new == 0 is a special case for destroying existing domains,
2683 * and it will not create the default domain.
2684 *
2685 * Call with hotplug lock and sched_domains_mutex held
2686 */
2687void partition_sched_domains_locked(int ndoms_new, cpumask_var_t doms_new[],
2688 struct sched_domain_attr *dattr_new)
2689{
2690 bool __maybe_unused has_eas = false;
2691 int i, j, n;
2692 int new_topology;
2693
2694 lockdep_assert_held(&sched_domains_mutex);
2695
2696 /* Let the architecture update CPU core mappings: */
2697 new_topology = arch_update_cpu_topology();
2698 /* Trigger rebuilding CPU capacity asymmetry data */
2699 if (new_topology)
2700 asym_cpu_capacity_scan();
2701
2702 if (!doms_new) {
2703 WARN_ON_ONCE(dattr_new);
2704 n = 0;
2705 doms_new = alloc_sched_domains(1);
2706 if (doms_new) {
2707 n = 1;
2708 cpumask_and(doms_new[0], cpu_active_mask,
2709 housekeeping_cpumask(HK_TYPE_DOMAIN));
2710 }
2711 } else {
2712 n = ndoms_new;
2713 }
2714
2715 /* Destroy deleted domains: */
2716 for (i = 0; i < ndoms_cur; i++) {
2717 for (j = 0; j < n && !new_topology; j++) {
2718 if (cpumask_equal(doms_cur[i], doms_new[j]) &&
2719 dattrs_equal(dattr_cur, i, dattr_new, j)) {
2720 struct root_domain *rd;
2721
2722 /*
2723 * This domain won't be destroyed and as such
2724 * its dl_bw->total_bw needs to be cleared. It
2725 * will be recomputed in function
2726 * update_tasks_root_domain().
2727 */
2728 rd = cpu_rq(cpumask_any(doms_cur[i]))->rd;
2729 dl_clear_root_domain(rd);
2730 goto match1;
2731 }
2732 }
2733 /* No match - a current sched domain not in new doms_new[] */
2734 detach_destroy_domains(doms_cur[i]);
2735match1:
2736 ;
2737 }
2738
2739 n = ndoms_cur;
2740 if (!doms_new) {
2741 n = 0;
2742 doms_new = &fallback_doms;
2743 cpumask_and(doms_new[0], cpu_active_mask,
2744 housekeeping_cpumask(HK_TYPE_DOMAIN));
2745 }
2746
2747 /* Build new domains: */
2748 for (i = 0; i < ndoms_new; i++) {
2749 for (j = 0; j < n && !new_topology; j++) {
2750 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2751 dattrs_equal(dattr_new, i, dattr_cur, j))
2752 goto match2;
2753 }
2754 /* No match - add a new doms_new */
2755 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
2756match2:
2757 ;
2758 }
2759
2760#if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
2761 /* Build perf domains: */
2762 for (i = 0; i < ndoms_new; i++) {
2763 for (j = 0; j < n && !sched_energy_update; j++) {
2764 if (cpumask_equal(doms_new[i], doms_cur[j]) &&
2765 cpu_rq(cpumask_first(doms_cur[j]))->rd->pd) {
2766 has_eas = true;
2767 goto match3;
2768 }
2769 }
2770 /* No match - add perf domains for a new rd */
2771 has_eas |= build_perf_domains(doms_new[i]);
2772match3:
2773 ;
2774 }
2775 sched_energy_set(has_eas);
2776#endif
2777
2778 /* Remember the new sched domains: */
2779 if (doms_cur != &fallback_doms)
2780 free_sched_domains(doms_cur, ndoms_cur);
2781
2782 kfree(dattr_cur);
2783 doms_cur = doms_new;
2784 dattr_cur = dattr_new;
2785 ndoms_cur = ndoms_new;
2786
2787 update_sched_domain_debugfs();
2788}
2789
2790/*
2791 * Call with hotplug lock held
2792 */
2793void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
2794 struct sched_domain_attr *dattr_new)
2795{
2796 mutex_lock(&sched_domains_mutex);
2797 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
2798 mutex_unlock(&sched_domains_mutex);
2799}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Scheduler topology setup/handling methods
4 */
5#include "sched.h"
6
7DEFINE_MUTEX(sched_domains_mutex);
8
9/* Protected by sched_domains_mutex: */
10cpumask_var_t sched_domains_tmpmask;
11cpumask_var_t sched_domains_tmpmask2;
12
13#ifdef CONFIG_SCHED_DEBUG
14
15static int __init sched_debug_setup(char *str)
16{
17 sched_debug_enabled = true;
18
19 return 0;
20}
21early_param("sched_debug", sched_debug_setup);
22
23static inline bool sched_debug(void)
24{
25 return sched_debug_enabled;
26}
27
28static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
29 struct cpumask *groupmask)
30{
31 struct sched_group *group = sd->groups;
32
33 cpumask_clear(groupmask);
34
35 printk(KERN_DEBUG "%*s domain-%d: ", level, "", level);
36
37 if (!(sd->flags & SD_LOAD_BALANCE)) {
38 printk("does not load-balance\n");
39 if (sd->parent)
40 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain has parent");
41 return -1;
42 }
43
44 printk(KERN_CONT "span=%*pbl level=%s\n",
45 cpumask_pr_args(sched_domain_span(sd)), sd->name);
46
47 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
48 printk(KERN_ERR "ERROR: domain->span does not contain CPU%d\n", cpu);
49 }
50 if (!cpumask_test_cpu(cpu, sched_group_span(group))) {
51 printk(KERN_ERR "ERROR: domain->groups does not contain CPU%d\n", cpu);
52 }
53
54 printk(KERN_DEBUG "%*s groups:", level + 1, "");
55 do {
56 if (!group) {
57 printk("\n");
58 printk(KERN_ERR "ERROR: group is NULL\n");
59 break;
60 }
61
62 if (!cpumask_weight(sched_group_span(group))) {
63 printk(KERN_CONT "\n");
64 printk(KERN_ERR "ERROR: empty group\n");
65 break;
66 }
67
68 if (!(sd->flags & SD_OVERLAP) &&
69 cpumask_intersects(groupmask, sched_group_span(group))) {
70 printk(KERN_CONT "\n");
71 printk(KERN_ERR "ERROR: repeated CPUs\n");
72 break;
73 }
74
75 cpumask_or(groupmask, groupmask, sched_group_span(group));
76
77 printk(KERN_CONT " %d:{ span=%*pbl",
78 group->sgc->id,
79 cpumask_pr_args(sched_group_span(group)));
80
81 if ((sd->flags & SD_OVERLAP) &&
82 !cpumask_equal(group_balance_mask(group), sched_group_span(group))) {
83 printk(KERN_CONT " mask=%*pbl",
84 cpumask_pr_args(group_balance_mask(group)));
85 }
86
87 if (group->sgc->capacity != SCHED_CAPACITY_SCALE)
88 printk(KERN_CONT " cap=%lu", group->sgc->capacity);
89
90 if (group == sd->groups && sd->child &&
91 !cpumask_equal(sched_domain_span(sd->child),
92 sched_group_span(group))) {
93 printk(KERN_ERR "ERROR: domain->groups does not match domain->child\n");
94 }
95
96 printk(KERN_CONT " }");
97
98 group = group->next;
99
100 if (group != sd->groups)
101 printk(KERN_CONT ",");
102
103 } while (group != sd->groups);
104 printk(KERN_CONT "\n");
105
106 if (!cpumask_equal(sched_domain_span(sd), groupmask))
107 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
108
109 if (sd->parent &&
110 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
111 printk(KERN_ERR "ERROR: parent span is not a superset of domain->span\n");
112 return 0;
113}
114
115static void sched_domain_debug(struct sched_domain *sd, int cpu)
116{
117 int level = 0;
118
119 if (!sched_debug_enabled)
120 return;
121
122 if (!sd) {
123 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
124 return;
125 }
126
127 printk(KERN_DEBUG "CPU%d attaching sched-domain(s):\n", cpu);
128
129 for (;;) {
130 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
131 break;
132 level++;
133 sd = sd->parent;
134 if (!sd)
135 break;
136 }
137}
138#else /* !CONFIG_SCHED_DEBUG */
139
140# define sched_debug_enabled 0
141# define sched_domain_debug(sd, cpu) do { } while (0)
142static inline bool sched_debug(void)
143{
144 return false;
145}
146#endif /* CONFIG_SCHED_DEBUG */
147
148static int sd_degenerate(struct sched_domain *sd)
149{
150 if (cpumask_weight(sched_domain_span(sd)) == 1)
151 return 1;
152
153 /* Following flags need at least 2 groups */
154 if (sd->flags & (SD_LOAD_BALANCE |
155 SD_BALANCE_NEWIDLE |
156 SD_BALANCE_FORK |
157 SD_BALANCE_EXEC |
158 SD_SHARE_CPUCAPACITY |
159 SD_ASYM_CPUCAPACITY |
160 SD_SHARE_PKG_RESOURCES |
161 SD_SHARE_POWERDOMAIN)) {
162 if (sd->groups != sd->groups->next)
163 return 0;
164 }
165
166 /* Following flags don't use groups */
167 if (sd->flags & (SD_WAKE_AFFINE))
168 return 0;
169
170 return 1;
171}
172
173static int
174sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
175{
176 unsigned long cflags = sd->flags, pflags = parent->flags;
177
178 if (sd_degenerate(parent))
179 return 1;
180
181 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
182 return 0;
183
184 /* Flags needing groups don't count if only 1 group in parent */
185 if (parent->groups == parent->groups->next) {
186 pflags &= ~(SD_LOAD_BALANCE |
187 SD_BALANCE_NEWIDLE |
188 SD_BALANCE_FORK |
189 SD_BALANCE_EXEC |
190 SD_ASYM_CPUCAPACITY |
191 SD_SHARE_CPUCAPACITY |
192 SD_SHARE_PKG_RESOURCES |
193 SD_PREFER_SIBLING |
194 SD_SHARE_POWERDOMAIN);
195 if (nr_node_ids == 1)
196 pflags &= ~SD_SERIALIZE;
197 }
198 if (~cflags & pflags)
199 return 0;
200
201 return 1;
202}
203
204static void free_rootdomain(struct rcu_head *rcu)
205{
206 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
207
208 cpupri_cleanup(&rd->cpupri);
209 cpudl_cleanup(&rd->cpudl);
210 free_cpumask_var(rd->dlo_mask);
211 free_cpumask_var(rd->rto_mask);
212 free_cpumask_var(rd->online);
213 free_cpumask_var(rd->span);
214 kfree(rd);
215}
216
217void rq_attach_root(struct rq *rq, struct root_domain *rd)
218{
219 struct root_domain *old_rd = NULL;
220 unsigned long flags;
221
222 raw_spin_lock_irqsave(&rq->lock, flags);
223
224 if (rq->rd) {
225 old_rd = rq->rd;
226
227 if (cpumask_test_cpu(rq->cpu, old_rd->online))
228 set_rq_offline(rq);
229
230 cpumask_clear_cpu(rq->cpu, old_rd->span);
231
232 /*
233 * If we dont want to free the old_rd yet then
234 * set old_rd to NULL to skip the freeing later
235 * in this function:
236 */
237 if (!atomic_dec_and_test(&old_rd->refcount))
238 old_rd = NULL;
239 }
240
241 atomic_inc(&rd->refcount);
242 rq->rd = rd;
243
244 cpumask_set_cpu(rq->cpu, rd->span);
245 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
246 set_rq_online(rq);
247
248 raw_spin_unlock_irqrestore(&rq->lock, flags);
249
250 if (old_rd)
251 call_rcu_sched(&old_rd->rcu, free_rootdomain);
252}
253
254void sched_get_rd(struct root_domain *rd)
255{
256 atomic_inc(&rd->refcount);
257}
258
259void sched_put_rd(struct root_domain *rd)
260{
261 if (!atomic_dec_and_test(&rd->refcount))
262 return;
263
264 call_rcu_sched(&rd->rcu, free_rootdomain);
265}
266
267static int init_rootdomain(struct root_domain *rd)
268{
269 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
270 goto out;
271 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
272 goto free_span;
273 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
274 goto free_online;
275 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
276 goto free_dlo_mask;
277
278#ifdef HAVE_RT_PUSH_IPI
279 rd->rto_cpu = -1;
280 raw_spin_lock_init(&rd->rto_lock);
281 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
282#endif
283
284 init_dl_bw(&rd->dl_bw);
285 if (cpudl_init(&rd->cpudl) != 0)
286 goto free_rto_mask;
287
288 if (cpupri_init(&rd->cpupri) != 0)
289 goto free_cpudl;
290 return 0;
291
292free_cpudl:
293 cpudl_cleanup(&rd->cpudl);
294free_rto_mask:
295 free_cpumask_var(rd->rto_mask);
296free_dlo_mask:
297 free_cpumask_var(rd->dlo_mask);
298free_online:
299 free_cpumask_var(rd->online);
300free_span:
301 free_cpumask_var(rd->span);
302out:
303 return -ENOMEM;
304}
305
306/*
307 * By default the system creates a single root-domain with all CPUs as
308 * members (mimicking the global state we have today).
309 */
310struct root_domain def_root_domain;
311
312void init_defrootdomain(void)
313{
314 init_rootdomain(&def_root_domain);
315
316 atomic_set(&def_root_domain.refcount, 1);
317}
318
319static struct root_domain *alloc_rootdomain(void)
320{
321 struct root_domain *rd;
322
323 rd = kzalloc(sizeof(*rd), GFP_KERNEL);
324 if (!rd)
325 return NULL;
326
327 if (init_rootdomain(rd) != 0) {
328 kfree(rd);
329 return NULL;
330 }
331
332 return rd;
333}
334
335static void free_sched_groups(struct sched_group *sg, int free_sgc)
336{
337 struct sched_group *tmp, *first;
338
339 if (!sg)
340 return;
341
342 first = sg;
343 do {
344 tmp = sg->next;
345
346 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
347 kfree(sg->sgc);
348
349 if (atomic_dec_and_test(&sg->ref))
350 kfree(sg);
351 sg = tmp;
352 } while (sg != first);
353}
354
355static void destroy_sched_domain(struct sched_domain *sd)
356{
357 /*
358 * A normal sched domain may have multiple group references, an
359 * overlapping domain, having private groups, only one. Iterate,
360 * dropping group/capacity references, freeing where none remain.
361 */
362 free_sched_groups(sd->groups, 1);
363
364 if (sd->shared && atomic_dec_and_test(&sd->shared->ref))
365 kfree(sd->shared);
366 kfree(sd);
367}
368
369static void destroy_sched_domains_rcu(struct rcu_head *rcu)
370{
371 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
372
373 while (sd) {
374 struct sched_domain *parent = sd->parent;
375 destroy_sched_domain(sd);
376 sd = parent;
377 }
378}
379
380static void destroy_sched_domains(struct sched_domain *sd)
381{
382 if (sd)
383 call_rcu(&sd->rcu, destroy_sched_domains_rcu);
384}
385
386/*
387 * Keep a special pointer to the highest sched_domain that has
388 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
389 * allows us to avoid some pointer chasing select_idle_sibling().
390 *
391 * Also keep a unique ID per domain (we use the first CPU number in
392 * the cpumask of the domain), this allows us to quickly tell if
393 * two CPUs are in the same cache domain, see cpus_share_cache().
394 */
395DEFINE_PER_CPU(struct sched_domain *, sd_llc);
396DEFINE_PER_CPU(int, sd_llc_size);
397DEFINE_PER_CPU(int, sd_llc_id);
398DEFINE_PER_CPU(struct sched_domain_shared *, sd_llc_shared);
399DEFINE_PER_CPU(struct sched_domain *, sd_numa);
400DEFINE_PER_CPU(struct sched_domain *, sd_asym);
401
402static void update_top_cache_domain(int cpu)
403{
404 struct sched_domain_shared *sds = NULL;
405 struct sched_domain *sd;
406 int id = cpu;
407 int size = 1;
408
409 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
410 if (sd) {
411 id = cpumask_first(sched_domain_span(sd));
412 size = cpumask_weight(sched_domain_span(sd));
413 sds = sd->shared;
414 }
415
416 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
417 per_cpu(sd_llc_size, cpu) = size;
418 per_cpu(sd_llc_id, cpu) = id;
419 rcu_assign_pointer(per_cpu(sd_llc_shared, cpu), sds);
420
421 sd = lowest_flag_domain(cpu, SD_NUMA);
422 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
423
424 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
425 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
426}
427
428/*
429 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
430 * hold the hotplug lock.
431 */
432static void
433cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
434{
435 struct rq *rq = cpu_rq(cpu);
436 struct sched_domain *tmp;
437
438 /* Remove the sched domains which do not contribute to scheduling. */
439 for (tmp = sd; tmp; ) {
440 struct sched_domain *parent = tmp->parent;
441 if (!parent)
442 break;
443
444 if (sd_parent_degenerate(tmp, parent)) {
445 tmp->parent = parent->parent;
446 if (parent->parent)
447 parent->parent->child = tmp;
448 /*
449 * Transfer SD_PREFER_SIBLING down in case of a
450 * degenerate parent; the spans match for this
451 * so the property transfers.
452 */
453 if (parent->flags & SD_PREFER_SIBLING)
454 tmp->flags |= SD_PREFER_SIBLING;
455 destroy_sched_domain(parent);
456 } else
457 tmp = tmp->parent;
458 }
459
460 if (sd && sd_degenerate(sd)) {
461 tmp = sd;
462 sd = sd->parent;
463 destroy_sched_domain(tmp);
464 if (sd)
465 sd->child = NULL;
466 }
467
468 sched_domain_debug(sd, cpu);
469
470 rq_attach_root(rq, rd);
471 tmp = rq->sd;
472 rcu_assign_pointer(rq->sd, sd);
473 dirty_sched_domain_sysctl(cpu);
474 destroy_sched_domains(tmp);
475
476 update_top_cache_domain(cpu);
477}
478
479struct s_data {
480 struct sched_domain ** __percpu sd;
481 struct root_domain *rd;
482};
483
484enum s_alloc {
485 sa_rootdomain,
486 sa_sd,
487 sa_sd_storage,
488 sa_none,
489};
490
491/*
492 * Return the canonical balance CPU for this group, this is the first CPU
493 * of this group that's also in the balance mask.
494 *
495 * The balance mask are all those CPUs that could actually end up at this
496 * group. See build_balance_mask().
497 *
498 * Also see should_we_balance().
499 */
500int group_balance_cpu(struct sched_group *sg)
501{
502 return cpumask_first(group_balance_mask(sg));
503}
504
505
506/*
507 * NUMA topology (first read the regular topology blurb below)
508 *
509 * Given a node-distance table, for example:
510 *
511 * node 0 1 2 3
512 * 0: 10 20 30 20
513 * 1: 20 10 20 30
514 * 2: 30 20 10 20
515 * 3: 20 30 20 10
516 *
517 * which represents a 4 node ring topology like:
518 *
519 * 0 ----- 1
520 * | |
521 * | |
522 * | |
523 * 3 ----- 2
524 *
525 * We want to construct domains and groups to represent this. The way we go
526 * about doing this is to build the domains on 'hops'. For each NUMA level we
527 * construct the mask of all nodes reachable in @level hops.
528 *
529 * For the above NUMA topology that gives 3 levels:
530 *
531 * NUMA-2 0-3 0-3 0-3 0-3
532 * groups: {0-1,3},{1-3} {0-2},{0,2-3} {1-3},{0-1,3} {0,2-3},{0-2}
533 *
534 * NUMA-1 0-1,3 0-2 1-3 0,2-3
535 * groups: {0},{1},{3} {0},{1},{2} {1},{2},{3} {0},{2},{3}
536 *
537 * NUMA-0 0 1 2 3
538 *
539 *
540 * As can be seen; things don't nicely line up as with the regular topology.
541 * When we iterate a domain in child domain chunks some nodes can be
542 * represented multiple times -- hence the "overlap" naming for this part of
543 * the topology.
544 *
545 * In order to minimize this overlap, we only build enough groups to cover the
546 * domain. For instance Node-0 NUMA-2 would only get groups: 0-1,3 and 1-3.
547 *
548 * Because:
549 *
550 * - the first group of each domain is its child domain; this
551 * gets us the first 0-1,3
552 * - the only uncovered node is 2, who's child domain is 1-3.
553 *
554 * However, because of the overlap, computing a unique CPU for each group is
555 * more complicated. Consider for instance the groups of NODE-1 NUMA-2, both
556 * groups include the CPUs of Node-0, while those CPUs would not in fact ever
557 * end up at those groups (they would end up in group: 0-1,3).
558 *
559 * To correct this we have to introduce the group balance mask. This mask
560 * will contain those CPUs in the group that can reach this group given the
561 * (child) domain tree.
562 *
563 * With this we can once again compute balance_cpu and sched_group_capacity
564 * relations.
565 *
566 * XXX include words on how balance_cpu is unique and therefore can be
567 * used for sched_group_capacity links.
568 *
569 *
570 * Another 'interesting' topology is:
571 *
572 * node 0 1 2 3
573 * 0: 10 20 20 30
574 * 1: 20 10 20 20
575 * 2: 20 20 10 20
576 * 3: 30 20 20 10
577 *
578 * Which looks a little like:
579 *
580 * 0 ----- 1
581 * | / |
582 * | / |
583 * | / |
584 * 2 ----- 3
585 *
586 * This topology is asymmetric, nodes 1,2 are fully connected, but nodes 0,3
587 * are not.
588 *
589 * This leads to a few particularly weird cases where the sched_domain's are
590 * not of the same number for each CPU. Consider:
591 *
592 * NUMA-2 0-3 0-3
593 * groups: {0-2},{1-3} {1-3},{0-2}
594 *
595 * NUMA-1 0-2 0-3 0-3 1-3
596 *
597 * NUMA-0 0 1 2 3
598 *
599 */
600
601
602/*
603 * Build the balance mask; it contains only those CPUs that can arrive at this
604 * group and should be considered to continue balancing.
605 *
606 * We do this during the group creation pass, therefore the group information
607 * isn't complete yet, however since each group represents a (child) domain we
608 * can fully construct this using the sched_domain bits (which are already
609 * complete).
610 */
611static void
612build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpumask *mask)
613{
614 const struct cpumask *sg_span = sched_group_span(sg);
615 struct sd_data *sdd = sd->private;
616 struct sched_domain *sibling;
617 int i;
618
619 cpumask_clear(mask);
620
621 for_each_cpu(i, sg_span) {
622 sibling = *per_cpu_ptr(sdd->sd, i);
623
624 /*
625 * Can happen in the asymmetric case, where these siblings are
626 * unused. The mask will not be empty because those CPUs that
627 * do have the top domain _should_ span the domain.
628 */
629 if (!sibling->child)
630 continue;
631
632 /* If we would not end up here, we can't continue from here */
633 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
634 continue;
635
636 cpumask_set_cpu(i, mask);
637 }
638
639 /* We must not have empty masks here */
640 WARN_ON_ONCE(cpumask_empty(mask));
641}
642
643/*
644 * XXX: This creates per-node group entries; since the load-balancer will
645 * immediately access remote memory to construct this group's load-balance
646 * statistics having the groups node local is of dubious benefit.
647 */
648static struct sched_group *
649build_group_from_child_sched_domain(struct sched_domain *sd, int cpu)
650{
651 struct sched_group *sg;
652 struct cpumask *sg_span;
653
654 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
655 GFP_KERNEL, cpu_to_node(cpu));
656
657 if (!sg)
658 return NULL;
659
660 sg_span = sched_group_span(sg);
661 if (sd->child)
662 cpumask_copy(sg_span, sched_domain_span(sd->child));
663 else
664 cpumask_copy(sg_span, sched_domain_span(sd));
665
666 atomic_inc(&sg->ref);
667 return sg;
668}
669
670static void init_overlap_sched_group(struct sched_domain *sd,
671 struct sched_group *sg)
672{
673 struct cpumask *mask = sched_domains_tmpmask2;
674 struct sd_data *sdd = sd->private;
675 struct cpumask *sg_span;
676 int cpu;
677
678 build_balance_mask(sd, sg, mask);
679 cpu = cpumask_first_and(sched_group_span(sg), mask);
680
681 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
682 if (atomic_inc_return(&sg->sgc->ref) == 1)
683 cpumask_copy(group_balance_mask(sg), mask);
684 else
685 WARN_ON_ONCE(!cpumask_equal(group_balance_mask(sg), mask));
686
687 /*
688 * Initialize sgc->capacity such that even if we mess up the
689 * domains and no possible iteration will get us here, we won't
690 * die on a /0 trap.
691 */
692 sg_span = sched_group_span(sg);
693 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
694 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
695}
696
697static int
698build_overlap_sched_groups(struct sched_domain *sd, int cpu)
699{
700 struct sched_group *first = NULL, *last = NULL, *sg;
701 const struct cpumask *span = sched_domain_span(sd);
702 struct cpumask *covered = sched_domains_tmpmask;
703 struct sd_data *sdd = sd->private;
704 struct sched_domain *sibling;
705 int i;
706
707 cpumask_clear(covered);
708
709 for_each_cpu_wrap(i, span, cpu) {
710 struct cpumask *sg_span;
711
712 if (cpumask_test_cpu(i, covered))
713 continue;
714
715 sibling = *per_cpu_ptr(sdd->sd, i);
716
717 /*
718 * Asymmetric node setups can result in situations where the
719 * domain tree is of unequal depth, make sure to skip domains
720 * that already cover the entire range.
721 *
722 * In that case build_sched_domains() will have terminated the
723 * iteration early and our sibling sd spans will be empty.
724 * Domains should always include the CPU they're built on, so
725 * check that.
726 */
727 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
728 continue;
729
730 sg = build_group_from_child_sched_domain(sibling, cpu);
731 if (!sg)
732 goto fail;
733
734 sg_span = sched_group_span(sg);
735 cpumask_or(covered, covered, sg_span);
736
737 init_overlap_sched_group(sd, sg);
738
739 if (!first)
740 first = sg;
741 if (last)
742 last->next = sg;
743 last = sg;
744 last->next = first;
745 }
746 sd->groups = first;
747
748 return 0;
749
750fail:
751 free_sched_groups(first, 0);
752
753 return -ENOMEM;
754}
755
756
757/*
758 * Package topology (also see the load-balance blurb in fair.c)
759 *
760 * The scheduler builds a tree structure to represent a number of important
761 * topology features. By default (default_topology[]) these include:
762 *
763 * - Simultaneous multithreading (SMT)
764 * - Multi-Core Cache (MC)
765 * - Package (DIE)
766 *
767 * Where the last one more or less denotes everything up to a NUMA node.
768 *
769 * The tree consists of 3 primary data structures:
770 *
771 * sched_domain -> sched_group -> sched_group_capacity
772 * ^ ^ ^ ^
773 * `-' `-'
774 *
775 * The sched_domains are per-CPU and have a two way link (parent & child) and
776 * denote the ever growing mask of CPUs belonging to that level of topology.
777 *
778 * Each sched_domain has a circular (double) linked list of sched_group's, each
779 * denoting the domains of the level below (or individual CPUs in case of the
780 * first domain level). The sched_group linked by a sched_domain includes the
781 * CPU of that sched_domain [*].
782 *
783 * Take for instance a 2 threaded, 2 core, 2 cache cluster part:
784 *
785 * CPU 0 1 2 3 4 5 6 7
786 *
787 * DIE [ ]
788 * MC [ ] [ ]
789 * SMT [ ] [ ] [ ] [ ]
790 *
791 * - or -
792 *
793 * DIE 0-7 0-7 0-7 0-7 0-7 0-7 0-7 0-7
794 * MC 0-3 0-3 0-3 0-3 4-7 4-7 4-7 4-7
795 * SMT 0-1 0-1 2-3 2-3 4-5 4-5 6-7 6-7
796 *
797 * CPU 0 1 2 3 4 5 6 7
798 *
799 * One way to think about it is: sched_domain moves you up and down among these
800 * topology levels, while sched_group moves you sideways through it, at child
801 * domain granularity.
802 *
803 * sched_group_capacity ensures each unique sched_group has shared storage.
804 *
805 * There are two related construction problems, both require a CPU that
806 * uniquely identify each group (for a given domain):
807 *
808 * - The first is the balance_cpu (see should_we_balance() and the
809 * load-balance blub in fair.c); for each group we only want 1 CPU to
810 * continue balancing at a higher domain.
811 *
812 * - The second is the sched_group_capacity; we want all identical groups
813 * to share a single sched_group_capacity.
814 *
815 * Since these topologies are exclusive by construction. That is, its
816 * impossible for an SMT thread to belong to multiple cores, and cores to
817 * be part of multiple caches. There is a very clear and unique location
818 * for each CPU in the hierarchy.
819 *
820 * Therefore computing a unique CPU for each group is trivial (the iteration
821 * mask is redundant and set all 1s; all CPUs in a group will end up at _that_
822 * group), we can simply pick the first CPU in each group.
823 *
824 *
825 * [*] in other words, the first group of each domain is its child domain.
826 */
827
828static struct sched_group *get_group(int cpu, struct sd_data *sdd)
829{
830 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
831 struct sched_domain *child = sd->child;
832 struct sched_group *sg;
833
834 if (child)
835 cpu = cpumask_first(sched_domain_span(child));
836
837 sg = *per_cpu_ptr(sdd->sg, cpu);
838 sg->sgc = *per_cpu_ptr(sdd->sgc, cpu);
839
840 /* For claim_allocations: */
841 atomic_inc(&sg->ref);
842 atomic_inc(&sg->sgc->ref);
843
844 if (child) {
845 cpumask_copy(sched_group_span(sg), sched_domain_span(child));
846 cpumask_copy(group_balance_mask(sg), sched_group_span(sg));
847 } else {
848 cpumask_set_cpu(cpu, sched_group_span(sg));
849 cpumask_set_cpu(cpu, group_balance_mask(sg));
850 }
851
852 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sched_group_span(sg));
853 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
854
855 return sg;
856}
857
858/*
859 * build_sched_groups will build a circular linked list of the groups
860 * covered by the given span, and will set each group's ->cpumask correctly,
861 * and ->cpu_capacity to 0.
862 *
863 * Assumes the sched_domain tree is fully constructed
864 */
865static int
866build_sched_groups(struct sched_domain *sd, int cpu)
867{
868 struct sched_group *first = NULL, *last = NULL;
869 struct sd_data *sdd = sd->private;
870 const struct cpumask *span = sched_domain_span(sd);
871 struct cpumask *covered;
872 int i;
873
874 lockdep_assert_held(&sched_domains_mutex);
875 covered = sched_domains_tmpmask;
876
877 cpumask_clear(covered);
878
879 for_each_cpu_wrap(i, span, cpu) {
880 struct sched_group *sg;
881
882 if (cpumask_test_cpu(i, covered))
883 continue;
884
885 sg = get_group(i, sdd);
886
887 cpumask_or(covered, covered, sched_group_span(sg));
888
889 if (!first)
890 first = sg;
891 if (last)
892 last->next = sg;
893 last = sg;
894 }
895 last->next = first;
896 sd->groups = first;
897
898 return 0;
899}
900
901/*
902 * Initialize sched groups cpu_capacity.
903 *
904 * cpu_capacity indicates the capacity of sched group, which is used while
905 * distributing the load between different sched groups in a sched domain.
906 * Typically cpu_capacity for all the groups in a sched domain will be same
907 * unless there are asymmetries in the topology. If there are asymmetries,
908 * group having more cpu_capacity will pickup more load compared to the
909 * group having less cpu_capacity.
910 */
911static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
912{
913 struct sched_group *sg = sd->groups;
914
915 WARN_ON(!sg);
916
917 do {
918 int cpu, max_cpu = -1;
919
920 sg->group_weight = cpumask_weight(sched_group_span(sg));
921
922 if (!(sd->flags & SD_ASYM_PACKING))
923 goto next;
924
925 for_each_cpu(cpu, sched_group_span(sg)) {
926 if (max_cpu < 0)
927 max_cpu = cpu;
928 else if (sched_asym_prefer(cpu, max_cpu))
929 max_cpu = cpu;
930 }
931 sg->asym_prefer_cpu = max_cpu;
932
933next:
934 sg = sg->next;
935 } while (sg != sd->groups);
936
937 if (cpu != group_balance_cpu(sg))
938 return;
939
940 update_group_capacity(sd, cpu);
941}
942
943/*
944 * Initializers for schedule domains
945 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
946 */
947
948static int default_relax_domain_level = -1;
949int sched_domain_level_max;
950
951static int __init setup_relax_domain_level(char *str)
952{
953 if (kstrtoint(str, 0, &default_relax_domain_level))
954 pr_warn("Unable to set relax_domain_level\n");
955
956 return 1;
957}
958__setup("relax_domain_level=", setup_relax_domain_level);
959
960static void set_domain_attribute(struct sched_domain *sd,
961 struct sched_domain_attr *attr)
962{
963 int request;
964
965 if (!attr || attr->relax_domain_level < 0) {
966 if (default_relax_domain_level < 0)
967 return;
968 else
969 request = default_relax_domain_level;
970 } else
971 request = attr->relax_domain_level;
972 if (request < sd->level) {
973 /* Turn off idle balance on this domain: */
974 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
975 } else {
976 /* Turn on idle balance on this domain: */
977 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
978 }
979}
980
981static void __sdt_free(const struct cpumask *cpu_map);
982static int __sdt_alloc(const struct cpumask *cpu_map);
983
984static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
985 const struct cpumask *cpu_map)
986{
987 switch (what) {
988 case sa_rootdomain:
989 if (!atomic_read(&d->rd->refcount))
990 free_rootdomain(&d->rd->rcu);
991 /* Fall through */
992 case sa_sd:
993 free_percpu(d->sd);
994 /* Fall through */
995 case sa_sd_storage:
996 __sdt_free(cpu_map);
997 /* Fall through */
998 case sa_none:
999 break;
1000 }
1001}
1002
1003static enum s_alloc
1004__visit_domain_allocation_hell(struct s_data *d, const struct cpumask *cpu_map)
1005{
1006 memset(d, 0, sizeof(*d));
1007
1008 if (__sdt_alloc(cpu_map))
1009 return sa_sd_storage;
1010 d->sd = alloc_percpu(struct sched_domain *);
1011 if (!d->sd)
1012 return sa_sd_storage;
1013 d->rd = alloc_rootdomain();
1014 if (!d->rd)
1015 return sa_sd;
1016
1017 return sa_rootdomain;
1018}
1019
1020/*
1021 * NULL the sd_data elements we've used to build the sched_domain and
1022 * sched_group structure so that the subsequent __free_domain_allocs()
1023 * will not free the data we're using.
1024 */
1025static void claim_allocations(int cpu, struct sched_domain *sd)
1026{
1027 struct sd_data *sdd = sd->private;
1028
1029 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
1030 *per_cpu_ptr(sdd->sd, cpu) = NULL;
1031
1032 if (atomic_read(&(*per_cpu_ptr(sdd->sds, cpu))->ref))
1033 *per_cpu_ptr(sdd->sds, cpu) = NULL;
1034
1035 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
1036 *per_cpu_ptr(sdd->sg, cpu) = NULL;
1037
1038 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
1039 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
1040}
1041
1042#ifdef CONFIG_NUMA
1043enum numa_topology_type sched_numa_topology_type;
1044
1045static int sched_domains_numa_levels;
1046static int sched_domains_curr_level;
1047
1048int sched_max_numa_distance;
1049static int *sched_domains_numa_distance;
1050static struct cpumask ***sched_domains_numa_masks;
1051#endif
1052
1053/*
1054 * SD_flags allowed in topology descriptions.
1055 *
1056 * These flags are purely descriptive of the topology and do not prescribe
1057 * behaviour. Behaviour is artificial and mapped in the below sd_init()
1058 * function:
1059 *
1060 * SD_SHARE_CPUCAPACITY - describes SMT topologies
1061 * SD_SHARE_PKG_RESOURCES - describes shared caches
1062 * SD_NUMA - describes NUMA topologies
1063 * SD_SHARE_POWERDOMAIN - describes shared power domain
1064 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
1065 *
1066 * Odd one out, which beside describing the topology has a quirk also
1067 * prescribes the desired behaviour that goes along with it:
1068 *
1069 * SD_ASYM_PACKING - describes SMT quirks
1070 */
1071#define TOPOLOGY_SD_FLAGS \
1072 (SD_SHARE_CPUCAPACITY | \
1073 SD_SHARE_PKG_RESOURCES | \
1074 SD_NUMA | \
1075 SD_ASYM_PACKING | \
1076 SD_ASYM_CPUCAPACITY | \
1077 SD_SHARE_POWERDOMAIN)
1078
1079static struct sched_domain *
1080sd_init(struct sched_domain_topology_level *tl,
1081 const struct cpumask *cpu_map,
1082 struct sched_domain *child, int cpu)
1083{
1084 struct sd_data *sdd = &tl->data;
1085 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
1086 int sd_id, sd_weight, sd_flags = 0;
1087
1088#ifdef CONFIG_NUMA
1089 /*
1090 * Ugly hack to pass state to sd_numa_mask()...
1091 */
1092 sched_domains_curr_level = tl->numa_level;
1093#endif
1094
1095 sd_weight = cpumask_weight(tl->mask(cpu));
1096
1097 if (tl->sd_flags)
1098 sd_flags = (*tl->sd_flags)();
1099 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
1100 "wrong sd_flags in topology description\n"))
1101 sd_flags &= ~TOPOLOGY_SD_FLAGS;
1102
1103 *sd = (struct sched_domain){
1104 .min_interval = sd_weight,
1105 .max_interval = 2*sd_weight,
1106 .busy_factor = 32,
1107 .imbalance_pct = 125,
1108
1109 .cache_nice_tries = 0,
1110 .busy_idx = 0,
1111 .idle_idx = 0,
1112 .newidle_idx = 0,
1113 .wake_idx = 0,
1114 .forkexec_idx = 0,
1115
1116 .flags = 1*SD_LOAD_BALANCE
1117 | 1*SD_BALANCE_NEWIDLE
1118 | 1*SD_BALANCE_EXEC
1119 | 1*SD_BALANCE_FORK
1120 | 0*SD_BALANCE_WAKE
1121 | 1*SD_WAKE_AFFINE
1122 | 0*SD_SHARE_CPUCAPACITY
1123 | 0*SD_SHARE_PKG_RESOURCES
1124 | 0*SD_SERIALIZE
1125 | 0*SD_PREFER_SIBLING
1126 | 0*SD_NUMA
1127 | sd_flags
1128 ,
1129
1130 .last_balance = jiffies,
1131 .balance_interval = sd_weight,
1132 .smt_gain = 0,
1133 .max_newidle_lb_cost = 0,
1134 .next_decay_max_lb_cost = jiffies,
1135 .child = child,
1136#ifdef CONFIG_SCHED_DEBUG
1137 .name = tl->name,
1138#endif
1139 };
1140
1141 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
1142 sd_id = cpumask_first(sched_domain_span(sd));
1143
1144 /*
1145 * Convert topological properties into behaviour.
1146 */
1147
1148 if (sd->flags & SD_ASYM_CPUCAPACITY) {
1149 struct sched_domain *t = sd;
1150
1151 for_each_lower_domain(t)
1152 t->flags |= SD_BALANCE_WAKE;
1153 }
1154
1155 if (sd->flags & SD_SHARE_CPUCAPACITY) {
1156 sd->flags |= SD_PREFER_SIBLING;
1157 sd->imbalance_pct = 110;
1158 sd->smt_gain = 1178; /* ~15% */
1159
1160 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1161 sd->flags |= SD_PREFER_SIBLING;
1162 sd->imbalance_pct = 117;
1163 sd->cache_nice_tries = 1;
1164 sd->busy_idx = 2;
1165
1166#ifdef CONFIG_NUMA
1167 } else if (sd->flags & SD_NUMA) {
1168 sd->cache_nice_tries = 2;
1169 sd->busy_idx = 3;
1170 sd->idle_idx = 2;
1171
1172 sd->flags |= SD_SERIALIZE;
1173 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
1174 sd->flags &= ~(SD_BALANCE_EXEC |
1175 SD_BALANCE_FORK |
1176 SD_WAKE_AFFINE);
1177 }
1178
1179#endif
1180 } else {
1181 sd->flags |= SD_PREFER_SIBLING;
1182 sd->cache_nice_tries = 1;
1183 sd->busy_idx = 2;
1184 sd->idle_idx = 1;
1185 }
1186
1187 /*
1188 * For all levels sharing cache; connect a sched_domain_shared
1189 * instance.
1190 */
1191 if (sd->flags & SD_SHARE_PKG_RESOURCES) {
1192 sd->shared = *per_cpu_ptr(sdd->sds, sd_id);
1193 atomic_inc(&sd->shared->ref);
1194 atomic_set(&sd->shared->nr_busy_cpus, sd_weight);
1195 }
1196
1197 sd->private = sdd;
1198
1199 return sd;
1200}
1201
1202/*
1203 * Topology list, bottom-up.
1204 */
1205static struct sched_domain_topology_level default_topology[] = {
1206#ifdef CONFIG_SCHED_SMT
1207 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
1208#endif
1209#ifdef CONFIG_SCHED_MC
1210 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
1211#endif
1212 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
1213 { NULL, },
1214};
1215
1216static struct sched_domain_topology_level *sched_domain_topology =
1217 default_topology;
1218
1219#define for_each_sd_topology(tl) \
1220 for (tl = sched_domain_topology; tl->mask; tl++)
1221
1222void set_sched_topology(struct sched_domain_topology_level *tl)
1223{
1224 if (WARN_ON_ONCE(sched_smp_initialized))
1225 return;
1226
1227 sched_domain_topology = tl;
1228}
1229
1230#ifdef CONFIG_NUMA
1231
1232static const struct cpumask *sd_numa_mask(int cpu)
1233{
1234 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
1235}
1236
1237static void sched_numa_warn(const char *str)
1238{
1239 static int done = false;
1240 int i,j;
1241
1242 if (done)
1243 return;
1244
1245 done = true;
1246
1247 printk(KERN_WARNING "ERROR: %s\n\n", str);
1248
1249 for (i = 0; i < nr_node_ids; i++) {
1250 printk(KERN_WARNING " ");
1251 for (j = 0; j < nr_node_ids; j++)
1252 printk(KERN_CONT "%02d ", node_distance(i,j));
1253 printk(KERN_CONT "\n");
1254 }
1255 printk(KERN_WARNING "\n");
1256}
1257
1258bool find_numa_distance(int distance)
1259{
1260 int i;
1261
1262 if (distance == node_distance(0, 0))
1263 return true;
1264
1265 for (i = 0; i < sched_domains_numa_levels; i++) {
1266 if (sched_domains_numa_distance[i] == distance)
1267 return true;
1268 }
1269
1270 return false;
1271}
1272
1273/*
1274 * A system can have three types of NUMA topology:
1275 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
1276 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
1277 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
1278 *
1279 * The difference between a glueless mesh topology and a backplane
1280 * topology lies in whether communication between not directly
1281 * connected nodes goes through intermediary nodes (where programs
1282 * could run), or through backplane controllers. This affects
1283 * placement of programs.
1284 *
1285 * The type of topology can be discerned with the following tests:
1286 * - If the maximum distance between any nodes is 1 hop, the system
1287 * is directly connected.
1288 * - If for two nodes A and B, located N > 1 hops away from each other,
1289 * there is an intermediary node C, which is < N hops away from both
1290 * nodes A and B, the system is a glueless mesh.
1291 */
1292static void init_numa_topology_type(void)
1293{
1294 int a, b, c, n;
1295
1296 n = sched_max_numa_distance;
1297
1298 if (sched_domains_numa_levels <= 1) {
1299 sched_numa_topology_type = NUMA_DIRECT;
1300 return;
1301 }
1302
1303 for_each_online_node(a) {
1304 for_each_online_node(b) {
1305 /* Find two nodes furthest removed from each other. */
1306 if (node_distance(a, b) < n)
1307 continue;
1308
1309 /* Is there an intermediary node between a and b? */
1310 for_each_online_node(c) {
1311 if (node_distance(a, c) < n &&
1312 node_distance(b, c) < n) {
1313 sched_numa_topology_type =
1314 NUMA_GLUELESS_MESH;
1315 return;
1316 }
1317 }
1318
1319 sched_numa_topology_type = NUMA_BACKPLANE;
1320 return;
1321 }
1322 }
1323}
1324
1325void sched_init_numa(void)
1326{
1327 int next_distance, curr_distance = node_distance(0, 0);
1328 struct sched_domain_topology_level *tl;
1329 int level = 0;
1330 int i, j, k;
1331
1332 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
1333 if (!sched_domains_numa_distance)
1334 return;
1335
1336 /* Includes NUMA identity node at level 0. */
1337 sched_domains_numa_distance[level++] = curr_distance;
1338 sched_domains_numa_levels = level;
1339
1340 /*
1341 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
1342 * unique distances in the node_distance() table.
1343 *
1344 * Assumes node_distance(0,j) includes all distances in
1345 * node_distance(i,j) in order to avoid cubic time.
1346 */
1347 next_distance = curr_distance;
1348 for (i = 0; i < nr_node_ids; i++) {
1349 for (j = 0; j < nr_node_ids; j++) {
1350 for (k = 0; k < nr_node_ids; k++) {
1351 int distance = node_distance(i, k);
1352
1353 if (distance > curr_distance &&
1354 (distance < next_distance ||
1355 next_distance == curr_distance))
1356 next_distance = distance;
1357
1358 /*
1359 * While not a strong assumption it would be nice to know
1360 * about cases where if node A is connected to B, B is not
1361 * equally connected to A.
1362 */
1363 if (sched_debug() && node_distance(k, i) != distance)
1364 sched_numa_warn("Node-distance not symmetric");
1365
1366 if (sched_debug() && i && !find_numa_distance(distance))
1367 sched_numa_warn("Node-0 not representative");
1368 }
1369 if (next_distance != curr_distance) {
1370 sched_domains_numa_distance[level++] = next_distance;
1371 sched_domains_numa_levels = level;
1372 curr_distance = next_distance;
1373 } else break;
1374 }
1375
1376 /*
1377 * In case of sched_debug() we verify the above assumption.
1378 */
1379 if (!sched_debug())
1380 break;
1381 }
1382
1383 if (!level)
1384 return;
1385
1386 /*
1387 * 'level' contains the number of unique distances
1388 *
1389 * The sched_domains_numa_distance[] array includes the actual distance
1390 * numbers.
1391 */
1392
1393 /*
1394 * Here, we should temporarily reset sched_domains_numa_levels to 0.
1395 * If it fails to allocate memory for array sched_domains_numa_masks[][],
1396 * the array will contain less then 'level' members. This could be
1397 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
1398 * in other functions.
1399 *
1400 * We reset it to 'level' at the end of this function.
1401 */
1402 sched_domains_numa_levels = 0;
1403
1404 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
1405 if (!sched_domains_numa_masks)
1406 return;
1407
1408 /*
1409 * Now for each level, construct a mask per node which contains all
1410 * CPUs of nodes that are that many hops away from us.
1411 */
1412 for (i = 0; i < level; i++) {
1413 sched_domains_numa_masks[i] =
1414 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
1415 if (!sched_domains_numa_masks[i])
1416 return;
1417
1418 for (j = 0; j < nr_node_ids; j++) {
1419 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
1420 if (!mask)
1421 return;
1422
1423 sched_domains_numa_masks[i][j] = mask;
1424
1425 for_each_node(k) {
1426 if (node_distance(j, k) > sched_domains_numa_distance[i])
1427 continue;
1428
1429 cpumask_or(mask, mask, cpumask_of_node(k));
1430 }
1431 }
1432 }
1433
1434 /* Compute default topology size */
1435 for (i = 0; sched_domain_topology[i].mask; i++);
1436
1437 tl = kzalloc((i + level + 1) *
1438 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
1439 if (!tl)
1440 return;
1441
1442 /*
1443 * Copy the default topology bits..
1444 */
1445 for (i = 0; sched_domain_topology[i].mask; i++)
1446 tl[i] = sched_domain_topology[i];
1447
1448 /*
1449 * Add the NUMA identity distance, aka single NODE.
1450 */
1451 tl[i++] = (struct sched_domain_topology_level){
1452 .mask = sd_numa_mask,
1453 .numa_level = 0,
1454 SD_INIT_NAME(NODE)
1455 };
1456
1457 /*
1458 * .. and append 'j' levels of NUMA goodness.
1459 */
1460 for (j = 1; j < level; i++, j++) {
1461 tl[i] = (struct sched_domain_topology_level){
1462 .mask = sd_numa_mask,
1463 .sd_flags = cpu_numa_flags,
1464 .flags = SDTL_OVERLAP,
1465 .numa_level = j,
1466 SD_INIT_NAME(NUMA)
1467 };
1468 }
1469
1470 sched_domain_topology = tl;
1471
1472 sched_domains_numa_levels = level;
1473 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
1474
1475 init_numa_topology_type();
1476}
1477
1478void sched_domains_numa_masks_set(unsigned int cpu)
1479{
1480 int node = cpu_to_node(cpu);
1481 int i, j;
1482
1483 for (i = 0; i < sched_domains_numa_levels; i++) {
1484 for (j = 0; j < nr_node_ids; j++) {
1485 if (node_distance(j, node) <= sched_domains_numa_distance[i])
1486 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
1487 }
1488 }
1489}
1490
1491void sched_domains_numa_masks_clear(unsigned int cpu)
1492{
1493 int i, j;
1494
1495 for (i = 0; i < sched_domains_numa_levels; i++) {
1496 for (j = 0; j < nr_node_ids; j++)
1497 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
1498 }
1499}
1500
1501#endif /* CONFIG_NUMA */
1502
1503static int __sdt_alloc(const struct cpumask *cpu_map)
1504{
1505 struct sched_domain_topology_level *tl;
1506 int j;
1507
1508 for_each_sd_topology(tl) {
1509 struct sd_data *sdd = &tl->data;
1510
1511 sdd->sd = alloc_percpu(struct sched_domain *);
1512 if (!sdd->sd)
1513 return -ENOMEM;
1514
1515 sdd->sds = alloc_percpu(struct sched_domain_shared *);
1516 if (!sdd->sds)
1517 return -ENOMEM;
1518
1519 sdd->sg = alloc_percpu(struct sched_group *);
1520 if (!sdd->sg)
1521 return -ENOMEM;
1522
1523 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
1524 if (!sdd->sgc)
1525 return -ENOMEM;
1526
1527 for_each_cpu(j, cpu_map) {
1528 struct sched_domain *sd;
1529 struct sched_domain_shared *sds;
1530 struct sched_group *sg;
1531 struct sched_group_capacity *sgc;
1532
1533 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
1534 GFP_KERNEL, cpu_to_node(j));
1535 if (!sd)
1536 return -ENOMEM;
1537
1538 *per_cpu_ptr(sdd->sd, j) = sd;
1539
1540 sds = kzalloc_node(sizeof(struct sched_domain_shared),
1541 GFP_KERNEL, cpu_to_node(j));
1542 if (!sds)
1543 return -ENOMEM;
1544
1545 *per_cpu_ptr(sdd->sds, j) = sds;
1546
1547 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
1548 GFP_KERNEL, cpu_to_node(j));
1549 if (!sg)
1550 return -ENOMEM;
1551
1552 sg->next = sg;
1553
1554 *per_cpu_ptr(sdd->sg, j) = sg;
1555
1556 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
1557 GFP_KERNEL, cpu_to_node(j));
1558 if (!sgc)
1559 return -ENOMEM;
1560
1561#ifdef CONFIG_SCHED_DEBUG
1562 sgc->id = j;
1563#endif
1564
1565 *per_cpu_ptr(sdd->sgc, j) = sgc;
1566 }
1567 }
1568
1569 return 0;
1570}
1571
1572static void __sdt_free(const struct cpumask *cpu_map)
1573{
1574 struct sched_domain_topology_level *tl;
1575 int j;
1576
1577 for_each_sd_topology(tl) {
1578 struct sd_data *sdd = &tl->data;
1579
1580 for_each_cpu(j, cpu_map) {
1581 struct sched_domain *sd;
1582
1583 if (sdd->sd) {
1584 sd = *per_cpu_ptr(sdd->sd, j);
1585 if (sd && (sd->flags & SD_OVERLAP))
1586 free_sched_groups(sd->groups, 0);
1587 kfree(*per_cpu_ptr(sdd->sd, j));
1588 }
1589
1590 if (sdd->sds)
1591 kfree(*per_cpu_ptr(sdd->sds, j));
1592 if (sdd->sg)
1593 kfree(*per_cpu_ptr(sdd->sg, j));
1594 if (sdd->sgc)
1595 kfree(*per_cpu_ptr(sdd->sgc, j));
1596 }
1597 free_percpu(sdd->sd);
1598 sdd->sd = NULL;
1599 free_percpu(sdd->sds);
1600 sdd->sds = NULL;
1601 free_percpu(sdd->sg);
1602 sdd->sg = NULL;
1603 free_percpu(sdd->sgc);
1604 sdd->sgc = NULL;
1605 }
1606}
1607
1608static struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
1609 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
1610 struct sched_domain *child, int cpu)
1611{
1612 struct sched_domain *sd = sd_init(tl, cpu_map, child, cpu);
1613
1614 if (child) {
1615 sd->level = child->level + 1;
1616 sched_domain_level_max = max(sched_domain_level_max, sd->level);
1617 child->parent = sd;
1618
1619 if (!cpumask_subset(sched_domain_span(child),
1620 sched_domain_span(sd))) {
1621 pr_err("BUG: arch topology borken\n");
1622#ifdef CONFIG_SCHED_DEBUG
1623 pr_err(" the %s domain not a subset of the %s domain\n",
1624 child->name, sd->name);
1625#endif
1626 /* Fixup, ensure @sd has at least @child CPUs. */
1627 cpumask_or(sched_domain_span(sd),
1628 sched_domain_span(sd),
1629 sched_domain_span(child));
1630 }
1631
1632 }
1633 set_domain_attribute(sd, attr);
1634
1635 return sd;
1636}
1637
1638/*
1639 * Build sched domains for a given set of CPUs and attach the sched domains
1640 * to the individual CPUs
1641 */
1642static int
1643build_sched_domains(const struct cpumask *cpu_map, struct sched_domain_attr *attr)
1644{
1645 enum s_alloc alloc_state;
1646 struct sched_domain *sd;
1647 struct s_data d;
1648 struct rq *rq = NULL;
1649 int i, ret = -ENOMEM;
1650
1651 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
1652 if (alloc_state != sa_rootdomain)
1653 goto error;
1654
1655 /* Set up domains for CPUs specified by the cpu_map: */
1656 for_each_cpu(i, cpu_map) {
1657 struct sched_domain_topology_level *tl;
1658
1659 sd = NULL;
1660 for_each_sd_topology(tl) {
1661 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
1662 if (tl == sched_domain_topology)
1663 *per_cpu_ptr(d.sd, i) = sd;
1664 if (tl->flags & SDTL_OVERLAP)
1665 sd->flags |= SD_OVERLAP;
1666 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
1667 break;
1668 }
1669 }
1670
1671 /* Build the groups for the domains */
1672 for_each_cpu(i, cpu_map) {
1673 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1674 sd->span_weight = cpumask_weight(sched_domain_span(sd));
1675 if (sd->flags & SD_OVERLAP) {
1676 if (build_overlap_sched_groups(sd, i))
1677 goto error;
1678 } else {
1679 if (build_sched_groups(sd, i))
1680 goto error;
1681 }
1682 }
1683 }
1684
1685 /* Calculate CPU capacity for physical packages and nodes */
1686 for (i = nr_cpumask_bits-1; i >= 0; i--) {
1687 if (!cpumask_test_cpu(i, cpu_map))
1688 continue;
1689
1690 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
1691 claim_allocations(i, sd);
1692 init_sched_groups_capacity(i, sd);
1693 }
1694 }
1695
1696 /* Attach the domains */
1697 rcu_read_lock();
1698 for_each_cpu(i, cpu_map) {
1699 rq = cpu_rq(i);
1700 sd = *per_cpu_ptr(d.sd, i);
1701
1702 /* Use READ_ONCE()/WRITE_ONCE() to avoid load/store tearing: */
1703 if (rq->cpu_capacity_orig > READ_ONCE(d.rd->max_cpu_capacity))
1704 WRITE_ONCE(d.rd->max_cpu_capacity, rq->cpu_capacity_orig);
1705
1706 cpu_attach_domain(sd, d.rd, i);
1707 }
1708 rcu_read_unlock();
1709
1710 if (rq && sched_debug_enabled) {
1711 pr_info("root domain span: %*pbl (max cpu_capacity = %lu)\n",
1712 cpumask_pr_args(cpu_map), rq->rd->max_cpu_capacity);
1713 }
1714
1715 ret = 0;
1716error:
1717 __free_domain_allocs(&d, alloc_state, cpu_map);
1718
1719 return ret;
1720}
1721
1722/* Current sched domains: */
1723static cpumask_var_t *doms_cur;
1724
1725/* Number of sched domains in 'doms_cur': */
1726static int ndoms_cur;
1727
1728/* Attribues of custom domains in 'doms_cur' */
1729static struct sched_domain_attr *dattr_cur;
1730
1731/*
1732 * Special case: If a kmalloc() of a doms_cur partition (array of
1733 * cpumask) fails, then fallback to a single sched domain,
1734 * as determined by the single cpumask fallback_doms.
1735 */
1736static cpumask_var_t fallback_doms;
1737
1738/*
1739 * arch_update_cpu_topology lets virtualized architectures update the
1740 * CPU core maps. It is supposed to return 1 if the topology changed
1741 * or 0 if it stayed the same.
1742 */
1743int __weak arch_update_cpu_topology(void)
1744{
1745 return 0;
1746}
1747
1748cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
1749{
1750 int i;
1751 cpumask_var_t *doms;
1752
1753 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
1754 if (!doms)
1755 return NULL;
1756 for (i = 0; i < ndoms; i++) {
1757 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
1758 free_sched_domains(doms, i);
1759 return NULL;
1760 }
1761 }
1762 return doms;
1763}
1764
1765void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
1766{
1767 unsigned int i;
1768 for (i = 0; i < ndoms; i++)
1769 free_cpumask_var(doms[i]);
1770 kfree(doms);
1771}
1772
1773/*
1774 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
1775 * For now this just excludes isolated CPUs, but could be used to
1776 * exclude other special cases in the future.
1777 */
1778int sched_init_domains(const struct cpumask *cpu_map)
1779{
1780 int err;
1781
1782 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_KERNEL);
1783 zalloc_cpumask_var(&sched_domains_tmpmask2, GFP_KERNEL);
1784 zalloc_cpumask_var(&fallback_doms, GFP_KERNEL);
1785
1786 arch_update_cpu_topology();
1787 ndoms_cur = 1;
1788 doms_cur = alloc_sched_domains(ndoms_cur);
1789 if (!doms_cur)
1790 doms_cur = &fallback_doms;
1791 cpumask_and(doms_cur[0], cpu_map, housekeeping_cpumask(HK_FLAG_DOMAIN));
1792 err = build_sched_domains(doms_cur[0], NULL);
1793 register_sched_domain_sysctl();
1794
1795 return err;
1796}
1797
1798/*
1799 * Detach sched domains from a group of CPUs specified in cpu_map
1800 * These CPUs will now be attached to the NULL domain
1801 */
1802static void detach_destroy_domains(const struct cpumask *cpu_map)
1803{
1804 int i;
1805
1806 rcu_read_lock();
1807 for_each_cpu(i, cpu_map)
1808 cpu_attach_domain(NULL, &def_root_domain, i);
1809 rcu_read_unlock();
1810}
1811
1812/* handle null as "default" */
1813static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
1814 struct sched_domain_attr *new, int idx_new)
1815{
1816 struct sched_domain_attr tmp;
1817
1818 /* Fast path: */
1819 if (!new && !cur)
1820 return 1;
1821
1822 tmp = SD_ATTR_INIT;
1823
1824 return !memcmp(cur ? (cur + idx_cur) : &tmp,
1825 new ? (new + idx_new) : &tmp,
1826 sizeof(struct sched_domain_attr));
1827}
1828
1829/*
1830 * Partition sched domains as specified by the 'ndoms_new'
1831 * cpumasks in the array doms_new[] of cpumasks. This compares
1832 * doms_new[] to the current sched domain partitioning, doms_cur[].
1833 * It destroys each deleted domain and builds each new domain.
1834 *
1835 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
1836 * The masks don't intersect (don't overlap.) We should setup one
1837 * sched domain for each mask. CPUs not in any of the cpumasks will
1838 * not be load balanced. If the same cpumask appears both in the
1839 * current 'doms_cur' domains and in the new 'doms_new', we can leave
1840 * it as it is.
1841 *
1842 * The passed in 'doms_new' should be allocated using
1843 * alloc_sched_domains. This routine takes ownership of it and will
1844 * free_sched_domains it when done with it. If the caller failed the
1845 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
1846 * and partition_sched_domains() will fallback to the single partition
1847 * 'fallback_doms', it also forces the domains to be rebuilt.
1848 *
1849 * If doms_new == NULL it will be replaced with cpu_online_mask.
1850 * ndoms_new == 0 is a special case for destroying existing domains,
1851 * and it will not create the default domain.
1852 *
1853 * Call with hotplug lock held
1854 */
1855void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1856 struct sched_domain_attr *dattr_new)
1857{
1858 int i, j, n;
1859 int new_topology;
1860
1861 mutex_lock(&sched_domains_mutex);
1862
1863 /* Always unregister in case we don't destroy any domains: */
1864 unregister_sched_domain_sysctl();
1865
1866 /* Let the architecture update CPU core mappings: */
1867 new_topology = arch_update_cpu_topology();
1868
1869 if (!doms_new) {
1870 WARN_ON_ONCE(dattr_new);
1871 n = 0;
1872 doms_new = alloc_sched_domains(1);
1873 if (doms_new) {
1874 n = 1;
1875 cpumask_and(doms_new[0], cpu_active_mask,
1876 housekeeping_cpumask(HK_FLAG_DOMAIN));
1877 }
1878 } else {
1879 n = ndoms_new;
1880 }
1881
1882 /* Destroy deleted domains: */
1883 for (i = 0; i < ndoms_cur; i++) {
1884 for (j = 0; j < n && !new_topology; j++) {
1885 if (cpumask_equal(doms_cur[i], doms_new[j])
1886 && dattrs_equal(dattr_cur, i, dattr_new, j))
1887 goto match1;
1888 }
1889 /* No match - a current sched domain not in new doms_new[] */
1890 detach_destroy_domains(doms_cur[i]);
1891match1:
1892 ;
1893 }
1894
1895 n = ndoms_cur;
1896 if (!doms_new) {
1897 n = 0;
1898 doms_new = &fallback_doms;
1899 cpumask_and(doms_new[0], cpu_active_mask,
1900 housekeeping_cpumask(HK_FLAG_DOMAIN));
1901 }
1902
1903 /* Build new domains: */
1904 for (i = 0; i < ndoms_new; i++) {
1905 for (j = 0; j < n && !new_topology; j++) {
1906 if (cpumask_equal(doms_new[i], doms_cur[j])
1907 && dattrs_equal(dattr_new, i, dattr_cur, j))
1908 goto match2;
1909 }
1910 /* No match - add a new doms_new */
1911 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
1912match2:
1913 ;
1914 }
1915
1916 /* Remember the new sched domains: */
1917 if (doms_cur != &fallback_doms)
1918 free_sched_domains(doms_cur, ndoms_cur);
1919
1920 kfree(dattr_cur);
1921 doms_cur = doms_new;
1922 dattr_cur = dattr_new;
1923 ndoms_cur = ndoms_new;
1924
1925 register_sched_domain_sysctl();
1926
1927 mutex_unlock(&sched_domains_mutex);
1928}